Search

Monocular depth estimation is an important but challenging task. Although the performance has been improved by adopting various encoder-decoder architectures, the estimated depth maps lack structure details and clear edges due to simple repeated upsampling. To solve this problem, this paper presents the novel LapUNet (Laplacian U-shape networks), in which the encoder adopts ResNeXt101, and the decoder is constructed with the novel DLRU (dynamic Laplacian residual U-shape) module. The DLRU module based on the U-shape structure can supplement high-frequency features by fusing dynamic Laplacian residual into the process of upsampling, and the residual is dynamically learnable due to the addition of convolutional operation. Also, the ASPP (atrous spatial pyramid pooling) module is introduced to capture image context at multiple scales though multiple parallel atrous convolutional layers, and the depth map fusion module is used for combining high and low frequency features from depth maps with different spatial resolution. Experiments demonstrate that the proposed model with moderate model size is superior to other previous competitors on the KITTI and NYU Depth V2 datasets. Furthermore, 3D reconstruction and target ranging by utilizing the estimated depth maps prove the effectiveness of our proposed method. [ABSTRACT FROM AUTHOR]

Self-supervised monocular depth estimation presents a promising result, which utilizes image sequences instead of challenging-to-source ground truth for training. The framework of most current studies on self-supervised depth estimation is based on fully convolutional or transformer architectures, and there is little discussion on the hybrid architecture. In this paper, we proposed TAMDepth, a novel framework that can effectively capture the local and global features of image sequences by combining convolutional blocks and transformer blocks. TAMDepth adopts multi-scale feature fusion convolutional modules capture local details in shallow layers while transformer blocks build the global dependency in higher layers. Furthermore, to enhance the representation of architecture, we introduce an adapter modulation that injects the spatial prior to the transformer blocks through cross-attention, which improves the ability of modeling the scene. Experiments demonstrate that our model exhibits state-of-the-art performance on the KITTI dataset and also shows strong generalization performance on the Make3D dataset. Source code is available at https://github.com/deansaice/TAMDepth. [ABSTRACT FROM AUTHOR]

The existing deep estimation networks often overlook the issue of computational efficiency while pursuing high accuracy. This paper proposes a lightweight self-supervised network that combines convolutional neural networks (CNN) and Transformers as the feature extraction and encoding layers for images, enabling the network to capture both local geometric and global semantic features for depth estimation. First, depth-separable convolution is used to construct a dilated convolution residual module based on a shallow network to improve the shallow CNN feature extraction receptive field. In the transformer, a multidepth separable convolution head transposed attention module is proposed to reduce the computational burden of spatial self-attention. In the feedforward network, a two-step gating mechanism is proposed to improve the nonlinear representation ability of the feedforward network. Finally, the CNN and transformer are integrated to implement a depth estimation network with a local-global context interaction function. Compared with other lightweight models, this model has fewer model parameters and higher estimation accuracy. It also has better generalizability for different outdoor datasets. Additionally, the inference speed can reach 87 FPS, achieving better real-time performance and accounting for both inference speed and estimation accuracy. [ABSTRACT FROM AUTHOR]

Monocular depth estimation (MDE) aims to predict pixel-level dense depth maps from a single RGB image. Some recent approaches mainly rely on encoder–decoder architectures to capture and process multi-scale features. However, they usually exploit heavier network at the expense of computational costs to obtain high-quality depth maps. In this paper, we propose a novel enriched multi-path vision transformer feature interaction network with an encoder–decoder architecture, denoted as EDFIDepth , which seeks a balance between computational costs and performance rather than pursuing the highest accuracy or extremely lightweight models. Specifically, an encoder called MPViT-D, incorporating multi-path vision transformer and a deep convolution module, is introduced to extract diverse features with both fine and coarse details at the same feature level with fewer parameters. Subsequently, we propose a lightweight decoder comprising two effective modules to establish multi-scale feature interaction: an encoder–decoder cross-feature matching (ED-CFM) module and a channel-level feature fusion (CLFF) module. The ED-CFM module is to establish connections between encoder–decoder features through a dual-path structure, where a cross-attention mechanism is deployed to enhance the relevance of multi-scale complementary depth information. Meanwhile, the CLFF module utilizes a channel attention mechanism to further fuse crucial depth information within the channels, thereby improving the accuracy of depth estimation. Extensive experiments on the indoor dataset NYUv2 and the outdoor dataset KITTI demonstrate that our method can achieve comparable state-of-the-art (SOTA) results while significantly reducing the number of trainable parameters. Our codes and approach are available at https://github.com/Zhangmg123/EDFIDEpth. [ABSTRACT FROM AUTHOR]

Monocular Simultaneous Localization And Mapping (SLAM) suffers from scale drift, leading to tracking failure due to scale ambiguity. Deep learning has significantly advanced self-supervised monocular depth estimation, enabling scale drift reduction. Nonetheless, current self-supervised learning approaches fail to provide scale-consistent depth maps, estimate depth in dynamic environments, or perceive multi-scale information. In response to these limitations, this paper proposes Dyna-MSDepth, a novel method for estimating multi-scale, stable, and reliable depth maps in dynamic environments. Dyna-MSDepth incorporates multi-scale high-order spatial semantic interaction into self-supervised training. This integration enhances the model’s capacity to discern intricate texture nuances and distant depth cues. Dyna-MSDepth is evaluated on challenging dynamic datasets, including KITTI, TUM, BONN, and DDAD, employing rigorous qualitative evaluations and quantitative experiments. Furthermore, the accuracy of the depth maps estimated by Dyna-MSDepth is assessed in monocular SLAM. Extensive experiments confirm the superior multi-scale depth estimation capabilities of Dyna-MSDepth, highlighting its significant value in dynamic environments. Code is available at . [ABSTRACT FROM AUTHOR]

Monocular depth estimation is an open challenge due to the ill-posed nature of the problem at hand. Deep learning techniques proved capable of producing acceptable depth estimation accuracy but the lack of robust depth cues within RGB images severally limits their performance. Coded aperture-based methods using phase and amplitude masks encode strong depth cues within 2D images by means of depth-dependent Point Spread Functions (PSFs) at the price of a reduced image quality. In this paper, we propose a novel end-to-end learning approach for depth from diffracted rotation. A phase mask that produces a Rotating Point Spread Function (RPSF) as a function of defocus is jointly optimized with the weights of a depth estimation neural network. To this aim, we introduce a differentiable physical model of the aperture mask and exploit an accurate simulation of the camera imaging pipeline. Our approach requires a significantly less complex model and less training data, yet it outperforms existing methods for monocular depth estimation on indoor benchmarks. In addition, we address the image degradation problem by incorporating a non-blind and nonuniform image deblurring module to recover the sharp all-in-focus image from its blurred counterpart. [ABSTRACT FROM AUTHOR]

This article deals with ensuring and increasing the safety of mobile robotic systems in human-machine collaboration. The goal of the research was to design and implement an artificial intelligence application that recognizes obstacles, including humans, and increases safety. The resulting mobile Android application uses a MiDaS model to generate a depth map of the environment from the drone's camera to approximate the distance from all obstacles to avoid the drone's collision. Besides, this work introduced us to DJI Mobile SDK and neural network optimizations for their use on smartphones. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Remote Sensing is the property of Editorial Office of Journal of Remote Sensing & Science Publishing Co. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Abstract Monocular depth estimation is an important but challenging task. Although the performance has been improved by adopting various encoder-decoder architectures, the estimated depth maps lack structure details and clear edges due to simple repeated upsampling. To solve this problem, this paper presents the novel LapUNet (Laplacian U-shape networks), in which the encoder adopts ResNeXt101, and the decoder is constructed with the novel DLRU (dynamic Laplacian residual U-shape) module. The DLRU module based on the U-shape structure can supplement high-frequency features by fusing dynamic Laplacian residual into the process of upsampling, and the residual is dynamically learnable due to the addition of convolutional operation. Also, the ASPP (atrous spatial pyramid pooling) module is introduced to capture image context at multiple scales though multiple parallel atrous convolutional layers, and the depth map fusion module is used for combining high and low frequency features from depth maps with different spatial resolution. Experiments demonstrate that the proposed model with moderate model size is superior to other previous competitors on the KITTI and NYU Depth V2 datasets. Furthermore, 3D reconstruction and target ranging by utilizing the estimated depth maps prove the effectiveness of our proposed method.

Abstract The existing deep estimation networks often overlook the issue of computational efficiency while pursuing high accuracy. This paper proposes a lightweight self-supervised network that combines convolutional neural networks (CNN) and Transformers as the feature extraction and encoding layers for images, enabling the network to capture both local geometric and global semantic features for depth estimation. First, depth-separable convolution is used to construct a dilated convolution residual module based on a shallow network to improve the shallow CNN feature extraction receptive field. In the transformer, a multidepth separable convolution head transposed attention module is proposed to reduce the computational burden of spatial self-attention. In the feedforward network, a two-step gating mechanism is proposed to improve the nonlinear representation ability of the feedforward network. Finally, the CNN and transformer are integrated to implement a depth estimation network with a local-global context interaction function. Compared with other lightweight models, this model has fewer model parameters and higher estimation accuracy. It also has better generalizability for different outdoor datasets. Additionally, the inference speed can reach 87 FPS, achieving better real-time performance and accounting for both inference speed and estimation accuracy.

This article deals with ensuring and increasing the safety of mobile robotic systems in human-machine collaboration. The goal of the research was to design and implement an artificial intelligence application that recognizes obstacles, including humans, and increases safety. The resulting mobile Android application uses a MiDaS model to generate a depth map of the environment from the drone’s camera to approximate the distance from all obstacles to avoid the drone’s collision. Besides, this work introduced us to DJI Mobile SDK and neural network optimizations for their use on smartphones.

The emergence of large-scale high-quality datasets has stimulated the rapid development of deep learning in recent years. However, most computer vision tasks focus on the visual modality only, resulting in a huge imbalance in the number of annotated data for other modalities. While several multi-modal datasets have been made available, the majority of them are confined to only two modalities, serving a single specific computer vision task. To redress the data deficiency for multi-modal learning and applications, a new dataset named UniMod1K is presented in this work. UniMod1K involves three data modalities: vision, depth, and language. For the vision and depth modalities, the UniMod1K dataset contains 1050 RGB-D sequences, comprising a total of some 2.5 million frames. Regarding the language modality, the proposed dataset includes 1050 sentences describing the target object in each video. To demonstrate the advantages of training on a larger multi-modal dataset, such as UniMod1K, and to stimulate research enabled by the dataset, we address several multi-modal tasks, namely multi-modal object tracking and monocular depth estimation. To establish a performance baseline, we propose novel baseline methods for RGB-D object tracking, vision-language tracking and vision-depth-language tracking. Additionally, we conduct comprehensive experiments for each of these tasks. The results highlight the potential of the UniMod1K dataset to improve the performance of multi-modal approaches. The dataset and codes can be accessed at https://github.com/xuefeng-zhu5/UniMod1K. [ABSTRACT FROM AUTHOR]

Monocular depth estimation aims to infer a depth map from a single image. Although supervised learning-based methods have achieved remarkable performance, they generally rely on a large amount of labor-intensively annotated data. Self-supervised methods, on the other hand, do not require any annotation of ground-truth depth and have recently attracted increasing attention. In this work, we propose a self-supervised monocular depth estimation network via binocular geometric correlation learning. Specifically, considering the inter-view geometric correlation, a binocular cue prediction module is presented to generate the auxiliary vision cue for the self-supervised learning of monocular depth estimation. Then, to deal with the occlusion in depth estimation, an occlusion interference attenuated constraint is developed to guide the supervision of the network by inferring the occlusion region and producing paired occlusion masks. Experimental results on two popular benchmark datasets have demonstrated that the proposed network obtains competitive results compared to state-of-the-art self-supervised methods and achieves comparable results to some popular supervised methods. [ABSTRACT FROM AUTHOR]

Based on the assumption of photometric consistency, self-supervised monocular depth estimation has been widely studied due to the advantage of avoiding costly annotations. However, it is sensitive to noise, occlusion issues and photometric changes. To overcome these problems, we propose a multi-task model with a dual-attention-based cross-task feature fusion module (DCFFM). We simultaneously predict depth and semantic with a shared encoder and two separate decoders, aiming to improve depth estimation with the enhancement of semantic supervision information. In DCFFM, we fuse the cross-task features with both pixel-wise and channel-wise attention, which fully excavate and make good use of the helpful information from the other task mutually. We compute both of two attentions in a one-to-all manner to capture global information while limiting the rapid growth of computation. Furthermore, we propose a novel data augmentation method called data exchange & recovery (DE &R), which performs inter-batch data exchange in both vertical and horizontal direction so as to increase the diversity of input data. It encourages the network to explore more diversified cues for depth estimation and avoid overfitting. And essentially, the corresponding outputs are further recovered in order to keep the geometry relationship and ensure the correct calculation of photometric loss. Extensive experiments on the KITTI dataset and the NYU-Depth-v2 dataset demonstrate that our method is very effective and achieves better performance compared with other state-of-the-art works. [ABSTRACT FROM AUTHOR]

In recent years, there has been extensive research and application of unsupervised monocular depth estimation methods for intelligent vehicles. However, a major limitation of most existing approaches is their inability to predict absolute depth values in physical units, as they generally suffer from the scale problem. Furthermore, most research efforts have focused on ground vehicles, neglecting the potential application of these methods to unmanned aerial vehicles (UAVs). To address these gaps, this paper proposes a novel absolute depth estimation method specifically designed for flight scenes using a monocular vision sensor, in which a geometry-based scale recovery algorithm serves as a post-processing stage of relative depth estimation results with scale consistency. By exploiting the feature correspondence between successive images and using the pose data provided by equipped navigation sensors, the scale factor between relative and absolute scales is calculated according to a multi-view geometry model, and then absolute depth maps are generated by pixel-wise multiplication of relative depth maps with the scale factor. As a result, the unsupervised monocular depth estimation technology is extended from relative depth estimation in semi-structured scenes to absolute depth estimation in unstructured scenes. Experiments on the publicly available Mid-Air dataset and customized data demonstrate the effectiveness of our method in different cases and settings, as well as its robustness to navigation sensor noise. The proposed method only requires UAVs to be equipped with monocular camera and common navigation sensors, and the obtained absolute depth information can be directly used for downstream tasks, which is significant for this kind of vehicle that has rarely been explored in previous depth estimation studies. [ABSTRACT FROM AUTHOR]

Accurate 3D image recognition, critical for autonomous driving safety, is shifting from the LIDAR-based point cloud to camera-based depth estimation technologies driven by cost considerations and the point cloud's limitations in detecting distant small objects. This research aims to enhance MDE (Monocular Depth Estimation) using a single camera, offering extreme cost-effectiveness in acquiring 3D environmental data. In particular, this paper focuses on novel data augmentation methods designed to enhance the accuracy of MDE. Our research addresses the challenge of limited MDE data quantities by proposing the use of synthetic-based augmentation techniques: Mask, Mask-Scale, and CutFlip. The implementation of these synthetic-based data augmentation strategies has demonstrably enhanced the accuracy of MDE models by 4.0% compared to the original dataset. Furthermore, this study introduces the RMS (Real-time Monocular Depth Estimation configuration considering Resolution, Efficiency, and Latency) algorithm, designed for the optimization of neural networks to augment the performance of contemporary monocular depth estimation technologies through a three-step process. Initially, it selects a model based on minimum latency and REL criteria, followed by refining the model's accuracy using various data augmentation techniques and loss functions. Finally, the refined model is compressed using quantization and pruning techniques to minimize its size for efficient on-device real-time applications. Experimental results from implementing the RMS algorithm indicated that, within the required latency and size constraints, the IEBins model exhibited the most accurate REL (absolute RELative error) performance, achieving a 0.0480 REL. Furthermore, the data augmentation combination of the original dataset with Flip, Mask, and CutFlip, alongside the SigLoss loss function, displayed the best REL performance, with a score of 0.0461 . The network compression technique using FP16 was analyzed as the most effective, reducing the model size by 83.4% compared to the original while maintaining the least impact on REL performance and latency. Finally, the performance of the RMS algorithm was validated on the on-device autonomous driving platform, NVIDIA Jetson AGX Orin, through which optimal deployment strategies were derived for various applications and scenarios requiring autonomous driving technologies. [ABSTRACT FROM AUTHOR]

Self-supervised monocular depth estimation can exhibit excellent performance in static environments due to the multi-view consistency assumption during the training process. However, it is hard to maintain depth consistency in dynamic scenes when considering the occlusion problem caused by moving objects. For this reason, we propose a method of self-supervised self-distillation for monocular depth estimation (SS-MDE) in dynamic scenes, where a deep network with a multi-scale decoder and a lightweight pose network are designed to predict depth in a self-supervised manner via the disparity, motion information, and the association between two adjacent frames in the image sequence. Meanwhile, in order to improve the depth estimation accuracy of static areas, the pseudo-depth images generated by the LeReS network are used to provide the pseudo-supervision information, enhancing the effect of depth refinement in static areas. Furthermore, a forgetting factor is leveraged to alleviate the dependency on the pseudo-supervision. In addition, a teacher model is introduced to generate depth prior information, and a multi-view mask filter module is designed to implement feature extraction and noise filtering. This can enable the student model to better learn the deep structure of dynamic scenes, enhancing the generalization and robustness of the entire model in a self-distillation manner. Finally, on four public data datasets, the performance of the proposed SS-MDE method outperformed several state-of-the-art monocular depth estimation techniques, achieving an accuracy ( δ 1 ) of 89% while minimizing the error (AbsRel) by 0.102 in NYU-Depth V2 and achieving an accuracy ( δ 1 ) of 87% while minimizing the error (AbsRel) by 0.111 in KITTI. [ABSTRACT FROM AUTHOR]

In recent years, many researchers have exploited multiple depth estimation architectures to produce high-quality depth maps from a single image. For monocular depth estimation, abundant multiscale features can significantly improve the prediction accuracy. Furthermore, multilevel refinement of the depth map through the model can effectively enhance the overall quality of the depth map. Therefore, we propose an efficient and effective module called light densely connected atrous spatial pyramid (LightDASP), which is employed to extract multiscale information at denser and larger scales from different levels of encoded features without significantly increasing the model size. Next, we propose a hierarchical reconstruction strategy that generates more accurate depth maps by refining the depth maps generated in the previous stage after each decoding stage. Additionally, to provide spatial location information to the decoder, the edge map is incorporated into the generation of a more rational refinement map. The experimental results, conducted on benchmark datasets in both indoor and outdoor scenes, demonstrate that our approach achieves efficient and competitive performance compared to existing methods for monocular depth estimation. We strike a balance between performance and efficiency, resulting in a model with greater potential for practical application. The code is available at upon article acceptance. [ABSTRACT FROM AUTHOR]

Depth estimation represents a prevalent research focus within the realm of computer vision. Existing depth estimation methodologies utilizing LiDAR (Light Detection and Ranging) technology typically obtain sparse depth data and are associated with elevated hardware expenses. Multi-view image-matching techniques necessitate prior knowledge of camera intrinsic parameters and frequently encounter challenges such as depth inconsistency, loss of details, and the blurring of edges. To tackle these challenges, the present study introduces a monocular depth estimation approach based on an end-to-end convolutional neural network. Specifically, a DNET backbone has been developed, incorporating dilated convolution and feature fusion mechanisms within the network architecture. By integrating semantic information from various receptive fields and levels, the model's capacity for feature extraction is augmented, thereby enhancing its sensitivity to nuanced depth variations within the image. Furthermore, we introduce a loss function optimization algorithm specifically designed to address class imbalance, thereby enhancing the overall predictive accuracy of the model. Training and validation conducted on the NYU Depth-v2 (New York University Depth Dataset Version 2) and KITTI (Karlsruhe Institute of Technology and Toyota Technological Institute) datasets demonstrate that our approach outperforms other algorithms in terms of various evaluation metrics. [ABSTRACT FROM AUTHOR]

Although amount of works are focused on monocular depth estimation, these works mainly study on the RGB spectrum, which has a poor performance on the case of nighttime, low light environment and even zero light environment. The images of other spectrum provide an opportunity to obtain depth without an active projector source. In this paper, we design a three-step architecture to realize monocular depth estimation by fusing cross-spectral stereo information. In the first step, we employ Spectral Translation Network to tackle with the problem that different spectral images have huge appearance differences and propose a disparity reservation loss to reserve disparity when translating. In the second step, we use Monocular Estimation Network to predict disparity of the principal spectrum, which is used for test. In the third step, we retrain the Spectral Translation Network with a generative optimization loss to improve the quality of image translation. Experiments show that our method achieves preeminent performance and reaches real-time speed. [ABSTRACT FROM AUTHOR]

Recent studies on self-supervised monocular depth estimation have achieved promising results, which are mainly based on the joint optimization of depth and pose estimation via high-level photometric loss. However, how to learn the latent and beneficial task-specific geometry representation from videos is still far from being explored. To tackle this issue, we propose two novel schemes to learn more effective representation from monocular videos: (i) an Inter-task Attention Model (IAM) to learn the geometric correlation representation between the depth and pose learning networks to make structure and motion information mutually beneficial; (ii) a Spatial-Temporal Memory Module (STMM) to exploit long-range geometric context representation among consecutive frames both spatially and temporally. Systematic ablation studies are conducted to demonstrate the effectiveness of each component. Evaluations on KITTI show that our method outperforms current state-of-the-art techniques. [ABSTRACT FROM AUTHOR]

A fuzzy mask applied to pixel-wise dissimilarity weighting is proposed to improve the monocular depth estimation in this study. The parameters in the monocular depth estimation model are learned unsupervised through the image reconstruction of binocular images. The significant reconstructed dissimilarity, which is challenging to reduce, always occurs at pixels outside the binocular overlap. The fuzzy mask is designed based on the binocular overlap to adjust the weight of the dissimilarity for each pixel. More than 68% of pixels with significant dissimilarity outside binocular overlap are suppressed with weights less than 0.5. The model with the proposed fuzzy mask would focus on learning the depth estimation for pixels within binocular overlap. Experiments on the KITTI dataset show that the inference of the fuzzy mask only increases the training time of the model by less than 1%, while the number of pixels whose depth is accurately estimated enhances, and the monocular depth estimation also improves. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Graphics is the property of Journal of Graphics Editorial Office and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Abstract Monocular depth estimation has a wide range of applications in the field of autostereoscopic displays, while accuracy and robustness in complex scenes are still a challenge. In this paper, we propose a depth estimation network for autostereoscopic displays, which aims at improving the accuracy of monocular depth estimation by fusing Vision Transformer (ViT) and Convolutional Neural Network (CNN). Our approach feeds the input image as a sequence of visual features into the ViT module and utilizes its global perception capability to extract high-level semantic features of the image. The relationship between the losses is quantified by adding a weight correction module to improve robustness of the model. Experimental evaluation results on several public datasets show that AMENet exhibits higher accuracy and robustness than existing methods in different scenarios and complex conditions. In addition, a detailed experimental analysis was conducted to verify the effectiveness and stability of our method. The accuracy improvement on the KITTI dataset compared to the baseline method is 4.4%. In summary, AMENet is a promising depth estimation method with sufficient high robustness and accuracy for monocular depth estimation tasks.

Existing self-supervised models for monocular depth estimation suffer from issues such as discontinuity, blurred edges, and unclear contours, particularly for small objects. We propose a self-supervised monocular depth estimation network with multi-scale selective Transformer feature fusion. To preserve more detailed features, this paper constructs a multi-scale encoder to extract features and leverages the self-attention mechanism of Transformer to capture global contextual information, enabling better depth prediction for small objects. Additionally, the multi-scale selective fusion module (MSSF) is also proposed, which can make full use of multi-scale feature information in the decoding part and perform selective fusion step by step, which can effectively eliminate noise and retain local detail features to obtain a clear depth map with clear edges. Experimental evaluations on the KITTI dataset demonstrate that the proposed algorithm achieves an absolute relative error (Abs Rel) of 0.098 and an accuracy rate (d) of 0.983. The results indicate that the proposed algorithm not only estimates depth values with high accuracy but also predicts the continuous depth map with clear edges. [ABSTRACT FROM AUTHOR]

Estimating depth from a single RGB image has a wide range of applications, such as in robot navigation and autonomous driving. Currently, Convolutional Neural Networks based on encoder–decoder architecture are the most popular methods to estimate depth maps. However, convolutional operators have limitations in modeling large-scale dependence, often leading to inaccurate depth predictions at object edges. To address these issues, a new edge-enhanced dual-stream monocular depth estimation method is introduced in this paper. ResNet and Swin Transformer are combined to better extract global and local features, which benefits the estimation of the depth map. To better integrate the information from the two branches of the encoder and the shallow branch of the decoder, we designed a lightweight decoder based on the multi-head Cross-Attention Module. Furthermore, in order to improve the boundary clarity of objects in the depth map, a loss function with an additional penalty for depth estimation error on the edges of objects is presented. The results on three datasets, NYU Depth V2, KITTI, and SUN RGB-D, show that the method presented in this paper achieves better performance for monocular depth estimation. Additionally, it has good generalization capabilities for various scenarios and real-world images. [ABSTRACT FROM AUTHOR]

Monocular depth estimation attracts great attention from computer vision researchers for its convenience in acquiring environment depth information. Recently classification-based MDE methods show its promising performance and begin to act as an essential role in many multi-view applications such as reconstruction and 3D object detection. However, existed classification-based MDE models usually apply fixed depth range discretization strategy across a whole scene. This fixed depth range discretization leads to the imbalance of discretization scale among different depth ranges, resulting in the inexact depth range localization. In this article, to alleviate the imbalanced depth range discretization problem in classification-based monocular depth estimation (MDE) method we follow the coarse-to-fine principle and propose a novel depth range discretization method called depth post-discretization (DPD). Based on a coarse depth anchor roughly indicating the depth range, the DPD generates the depth range discretization adaptively for every position. The depth range discretization with DPD is more fine-grained around the actual depth, which is beneficial for locating the depth range more precisely for each scene position. Besides, to better manage the prediction of the coarse depth anchor and depth probability distribution for calculating the final depth, we design a dual-decoder transformer-based network, i.e., DPDFormer, which is more compatible with our proposed DPD method. We evaluate DPDFormer on popular depth datasets NYU Depth V2 and KITTI. The experimental results prove the superior performance of our proposed method. [ABSTRACT FROM AUTHOR]

Robust monocular depth estimation (MDE) aims at learning a unified model that works across diverse real-world scenes, which is an important and active topic in computer vision. In this paper, we present Megatron_RVC, our winning solution for the monocular depth challenge in the Robust Vision Challenge (RVC) 2022, where we tackle the challenging problem from three perspectives: network architecture, training strategy and dataset. In particular, we made three contributions towards robust MDE: (1) we built a neural network with high capacity to enable flexible and accurate monocular depth predictions, which contains dedicated components to provide content-aware embeddings and to improve the richness of the details; (2) we proposed a novel mixing training strategy to handle real-world images with different aspect ratios, resolutions and apply tailored loss functions based on the properties of their depth maps; (3) to train a unified network model that covers diverse real-world scenes, we used over 1 million images from different datasets. As of 3rd October 2022, our unified model ranked consistently first across three benchmarks (KITTI, MPI Sintel, and VIPER) among all participants. [ABSTRACT FROM AUTHOR]

Monocular depth estimation is a crucial task in computer vision, and self-supervised algorithms are gaining popularity due to their independence from expensive ground truth supervision. However, current self-supervised algorithms may not provide accurate estimation and may suffer from distorted boundaries when applied to indoor scenes. Combining multi-scale features is an important research direction in image segmentation to achieve accurate estimation and resolve boundary distortion. However, there are few studies on indoor self-supervised algorithms in this regard. To solve this issue, we propose a novel full-scale feature information fusion approach that includes a full-scale skip-connection and a full-scale feature fusion block. This approach can aggregate the high-level and low-level information of all scale feature maps during the network's encoding and decoding process to compensate for the network's loss of cross-layer feature information. The proposed full-scale feature fusion improves accuracy and reduces the decoder parameters. To fully exploit the superiority of the full-scale feature fusion module, we replace the encoder backbone from ResNet with the more advanced ResNeSt. Combining these two methods results in a significant improvement in prediction accuracy. We have extensively evaluated our approach on the indoor benchmark datasets NYU Depth V2 and ScanNet. Our experimental results demonstrate that our method outperforms existing algorithms, particularly on NYU Depth V2, where our precision is raised to 83.8%. [ABSTRACT FROM AUTHOR]

Depth estimation is an essential component of computer vision applications for environment perception, 3D reconstruction and scene understanding. Among the available methods, self-supervised monocular depth estimation is noteworthy for its cost-effectiveness, ease of installation and data accessibility. However, there are two challenges with current methods. Firstly, the scale factor of self-supervised monocular depth estimation is uncertain, which poses significant difficulties for practical applications. Secondly, the depth prediction accuracy for high-resolution images is still unsatisfactory, resulting in low utilization of computational resources. We propose a novel solution to address these challenges with three specific contributions. Firstly, an interleaved depth network skip-connection structure and a new depth network decoder are proposed to improve the depth prediction accuracy for high-resolution images. Secondly, a data vertical splicing module is suggested as a data enhancement method to obtain more non-vertical features and improve model generalization. Lastly, a scale recovery module is proposed to recover the accurate absolute depth without additional sensors, which solves the issue of uncertainty in the scale factor. The experimental results demonstrate that the proposed framework significantly improves the prediction accuracy of high-resolution images. In particular, the novel network structure and data vertical splicing module contribute significantly to this improvement. Moreover, in a scenario where the camera height is fixed and the ground is flat, the effect of scale recovery module is comparable to that achieved by using ground truth. Overall, the RSANet framework offers a promising solution to solve the existing challenges in self-supervised monocular depth estimation. [ABSTRACT FROM AUTHOR]

A lot of research has been conducted in recent years on stereo depth estimation techniques, taking the traditional approach to a new level such that it is in an appreciably good form for competing in the depth estimation market with other methods, despite its few demerits. Sufficient progress in accuracy and depth computation speed has manifested during the period. Over the years, stereo depth estimation has been provided with various training modes, such as supervised, self-supervised, and unsupervised, before deploying it for real-time performance. These modes are to be used depending on the application and/or the availability of datasets for training. Deep learning, on the other hand, has provided the stereo depth estimation methods with a new life to breathe in the form of enhanced accuracy and quality of images, attempting to successfully reduce the residual errors in stages in some of the methods. Furthermore, depth estimation from a single RGB image has been intricate since it is an ill-posed problem with a lack of geometric constraints and ambiguities. However, this monocular depth estimation has gained popularity in recent years due to the development in the field, with appreciable improvements in the accuracy of depth maps and optimization of computational time. The help is mostly due to the usage of CNNs (Convolutional Neural Networks) and other deep learning methods, which help augment the feature-extraction phenomenon for the process and enhance the quality of depth maps/accuracy of MDE (monocular depth estimation). Monocular depth estimation has seen improvements in many algorithms that can be deployed to give depth maps with better clarity and details around the edges and fine boundaries, which thus helps in delineating between thin structures. This paper reviews various recent deep learning-based stereo and monocular depth prediction techniques emphasizing the successes achieved so far, the challenges acquainted with them, and those that can be expected shortly. [ABSTRACT FROM AUTHOR]

3D reconstruction is a long-standing research topic in the photogrammetric and computer vision communities; although a plethora of open-source and commercial solutions for 3D reconstruction have been released in the last few years, several open challenges and limitations still exist. Undoubtedly, deep learning algorithms have demonstrated great potential in several remote sensing tasks, including image-based 3D reconstruction. State-of-the-art monocular and stereo algorithms leverage deep learning techniques and achieve increased performance in depth estimation and 3D reconstruction. However, one of the limitations of such methods is that they highly rely on large training sets that are often tedious to obtain; even when available, they typically refer to indoor, close-range scenarios and low-resolution images. Especially while considering UAV (Unmanned Aerial Vehicle) scenarios, such data are not available and domain adaptation is not a trivial challenge. To fill this gap, the UAV-based multi-sensor dataset for geospatial research (UseGeo - https://usegeo.fbk.eu/home) is introduced in this paper. It contains both image and LiDAR data and aims to support relevant research in photogrammetry and computer vision with a useful training set for both stereo and monocular 3D reconstruction algorithms. In this regard, the dataset provides ground truth data for both point clouds and depth maps. In addition, UseGeo can be also a valuable dataset for other tasks such as feature extraction and matching, aerial triangulation, or image and LiDAR co-registration. The paper introduces the UseGeo dataset and validates some state-of-the-art algorithms to assess their usability for both monocular and multi-view 3D reconstruction.

Calculating depth using just one image is a crucial issue since it has applications in numerous computer vision domains. Although some recent works directly obtain the depth map through some complex and powerful networks, we want to combine the encoder and decoder feature maps more effectively. To this end, we propose a novel U-Net like network. The encoder is based on Swin Transformer. For the decoder, we propose Polarization Cross Attention to effectively combine codec features by optimizing the initialization of the k and v vector. In order to conduct a more in-depth global analysis of the decoded output, a Sparse Transformer post-processing module is proposed. In the Sparse Transformer module, we adopt Kullback–Leibler divergence to obtain a sparse Q matrix and achieve O((hw)ln(hw)) in time complexity and memory usage. Results from experiments utilizing the KITTI and NYUV2 datasets demonstrate how well the suggested strategy enhances the precision of monocular depth perception when compared with state-of-the-art methods.

This article addresses the problem of constructing maps for 2D simulated environments. An algorithm based on monocular depth estimation is proposed achieving comparable accuracy to methods utilizing expensive sensors such as RGBD cameras and LIDARs. To solve the problem, we employ a multi-stage approach. First, a neural network predicts a relative disparity map from an RGB flow provided by RGBD camera. Using depth measurements from the same camera, two parameters are estimated that connect the relative and absolute displacement maps in the form of a linear regression relation. Based on a simpler RGB camera, by applying a neural network and estimates of scaling parameters, an estimate of the absolute displacement map is formed, which allows to obtain an estimate of the depth map. Thus, a virtual scanner has been designed providing Cartographer SLAM with depth information for environment mapping. The proposed algorithm was evaluated on a ROS 2.0 simulation of a simple mobile robot. It achieves faster depth prediction compared to other depth estimation algorithms. Furthermore, maps generated by our approach demonstrated a high overlap ratio with those obtained using an ideal RGBD camera. The proposed algorithm can find applicability in crucial tasks for mobile robots, like obstacle avoidance, and path planning. Moreover, it can be used to generate accurate cost maps, enhancing safety and adaptability in mobile robot navigation.

Non-Lambertian surfaces are special surfaces that can cause specific type of reflectances called specularities, which pose a potential issue in industrial SLAM. This article reviews fundamental surface reflectance models, modern state-of-the-art computer vision algorithms and two public datasets, KITTI and DiLiGenT, related to non-Lambertian surfaces' research. A new dataset, SPINS, is presented for the purpose of studying non-Lambertian surfaces in navigation and an empirical performance evaluation with ResNeXt-101-WSL, ORB SLAM 3 and TartanVO is performed on the data. The article concludes with discussion about the results of empirical evaluation and the findings of the survey.

To improve the prediction accuracy of monocular depth estimation networks and address issues such as edge blurring and excessive artifacts in the generated depth maps, this paper proposes a deep network architecture based on a global-local feature fusion module and a residual pooling module. The encoder utilizes a Hierarchical Transformer, while the decoder incorporates a U-Net structure model that combines multi-dimensional attention features aggregation and residual pooling. The residual pooling module facilitates better extraction of background contextual information from the feature maps to obtain more precise scene depth information. The global-local feature fusion module enables the network to learn features that encompass both global and local information. Experimental evaluations conducted on the NYU Depth V2 and KITTI datasets demonstrate that the proposed method achieves a $\delta 1$ of 0.916 on the NYU Depth V2 dataset, along with enhanced generalization ability and robustness. Furthermore, the effectiveness of each module is validated through ablation studies on the NYU Depth V2 dataset.

A lot of research has been conducted in recent years on stereo depth estimation techniques, taking the traditional approach to a new level such that it is in an appreciably good form for competing in the depth estimation market with other methods, despite its few demerits. Sufficient progress in accuracy and depth computation speed has manifested during the period. Over the years, stereo depth estimation has been provided with various training modes, such as supervised, self-supervised, and unsupervised, before deploying it for real-time performance. These modes are to be used depending on the application and/or the availability of datasets for training. Deep learning, on the other hand, has provided the stereo depth estimation methods with a new life to breathe in the form of enhanced accuracy and quality of images, attempting to successfully reduce the residual errors in stages in some of the methods. Furthermore, depth estimation from a single RGB image has been intricate since it is an ill-posed problem with a lack of geometric constraints and ambiguities. However, this monocular depth estimation has gained popularity in recent years due to the development in the field, with appreciable improvements in the accuracy of depth maps and optimization of computational time. The help is mostly due to the usage of CNNs (Convolutional Neural Networks) and other deep learning methods, which help augment the feature-extraction phenomenon for the process and enhance the quality of depth maps/accuracy of MDE (monocular depth estimation). Monocular depth estimation has seen improvements in many algorithms that can be deployed to give depth maps with better clarity and details around the edges and fine boundaries, which thus helps in delineating between thin structures. This paper reviews various recent deep learning-based stereo and monocular depth prediction techniques emphasizing the successes achieved so far, the challenges acquainted with them, and those that can be expected shortly.

Access to safe and secure food is an essential human requirement for a sustainable world. Ensuring the safety of fresh food products necessitates accurate analytical techniques. This study presents an innovative food safety approach using rapid 3D reconstruction and digital twin analysis in mixed reality. While 3D models offer comprehensive insights into fresh food, conventional reconstruction methods are often complex due to multiple cameras or sensors. To address this, the research employs monocular depth estimation for detailed 3D reconstruction from a single camera view. This architecture uses web service protocols and the Global-Local Path Networks (GLPN) model for monocular depth perception, effectively translating 2D images into intricate 3D structures. These digital twins establish a strong foundation for enhanced analysis of food products, facilitating a better understanding of safety risks. The study also explores the applicability of transferring 3D models to Microsoft HoloLens 2, enabling immersive visualization and novel avenues for analysis. These techniques hold significance for the food industry, spanning scene comprehension, precision agriculture, robotics, augmented reality, and medical imaging. The methodology includes assessing surface degradation, with a focus on bruises on fresh produce. Employing a mesh overlay technique on the resulting 3D models distinctly showcases the impact of bruising. Sequential data recordings effectively track bruise expansion, shedding light on structural compromises. This technique dynamically portrays bruise progression, thus enriching comprehension of its implications on produce quality.

Visual place recognition is one of the core modern computer vision tasks concerned with identifying location based on the image taken there. Modern state-of-the-art approaches heavily rely on RGB images which are largely affected by changes in the same scene such as varying daytime, illumination, seasonal changes, and presence of dynamic objects (people, vehicles). This results into a large difference between the images in the training dataset and the ones taken by a person in real life at the same place as a part of some application, rendering modern approaches less effective. To deal with this problem, we propose a novel approach that uses only geometrical information (shapes of buildings, terrains, trees, and their relevant positions) obtained from depth and semantic maps inpainted to remove dynamic objects. In this paper, we study two versions of the pipeline: the first one uses direct inpainting, and the second utilizes synthetic data to improve the inpainting process. Our most efficient model achieved 60.6% correct answers with synthetic refinement. With direct inpainting, it kept metrics high at 51.1%. With these compelling results, our approach offers a novel and effective alternative to known algorithms, making it an exciting avenue for future research in visual place recognition.

In this paper, we propose a novel method for monocular depth estimation using the hourglass neck module. The proposed method has the following originality. First, feature maps are extracted from Swin Transformer V2 using a masked image modeling (MIM) pretrained model. Since Swin Transformer V2 has a different patch size for each attention stage, it is easier to extract local and global features from images input by the vision transformer (ViT)-based encoder. Second, to maintain the polymorphism and local inductive bias of the feature map extracted from Swin Transformer V2, a feature map is input into the hourglass neck module. Third, deformable attention can be used at the waist of the hourglass neck module to reduce the computation cost and highlight the locality of the feature map. Finally, the feature map traverses the neck and proceeds through a decoder, comprised of a deconvolution layer and an upsampling layer, to generate a depth image. To evaluate the objective reliability of the proposed method in this paper, we used the NYU Depth V2 dataset to compare and evaluate the methods published in other papers. As a result of the experiment, the RMSE value of the novel method for monocular depth estimation using the hourglass neck module proposed in this paper was 0.274, which was lower than those published in other papers. The lower the RMSE value, the better the depth estimation method; therefore, its efficiency compared to other techniques has been proven. [ABSTRACT FROM AUTHOR]

Now the dense depth prediction by single image and a few sparse depth measurements has attracted more and more attention because it provides a low-cost and efficient solution for estimating high-quality depth information. But the current existing methods for the field only take sparse depth as an independent dimension, and the relationship between sparse depth and image itself is always ignored, which undoubtedly limits the improvement of prediction accuracy. For solving the problem, in this paper, a sparse depth densification method is proposed to fully mine the relationship between sparse depth and image for achieving more accurate depth estimation. Based on a priori that the object areas of same category have similar depth values, a Depth Densification Map (DDM) is constructed by the segmentation label obtained from unsupervised image segmentation and sparse depth to realize sparse depth densification. Meantime, considering the potential error of DDM, a Depth Error Map (DEM) is designed to further correct DDM. Then, we use the idea of multi-scale fusion to build a depth estimation network. Finally, the proposed maps are combined with single image as the input of the network, and used to carry out the actual training and testing. Extensive experiments on NYU Depth v2 and Make3D datasets demonstrate the superiority of our proposed approach. Our code is available at https://github.com/Jennifer108/Sparse-Depth-Densification. https://github.com/Jennifer108/Sparse-Depth-Densification. [ABSTRACT FROM AUTHOR]

Copyright of Systems Engineering & Electronics is the property of Journal of Systems Engineering & Electronics Editorial Department and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Monocular depth estimation is a task aimed at predicting pixel-level distances from a single RGB image. This task holds significance in various applications including autonomous driving and robotics. In particular, the recognition of surrounding environments is important to avoid collisions during autonomous parking. Fisheye cameras are adequate to acquire visual information from a wide field of view, reducing blind spots and preventing potential collisions. While there have been increasing demands for fisheye cameras in visual-recognition systems, existing research on depth estimation has primarily focused on pinhole camera images. Moreover, depth estimation from fisheye images poses additional challenges due to strong distortion and the lack of public datasets. In this work, we propose a novel underground parking lot dataset called JBNU-Depth360, which consists of fisheye camera images and their corresponding LiDAR projections. Our proposed dataset was composed of 4221 pairs of fisheye images and their corresponding LiDAR point clouds, which were obtained from six driving sequences. Furthermore, we employed a knowledge-distillation technique to improve the performance of the state-of-the-art depth-estimation models. The teacher–student learning framework allows the neural network to leverage the information in dense depth predictions and sparse LiDAR projections. Experiments were conducted on the KITTI-360 and JBNU-Depth360 datasets for analyzing the performance of existing depth-estimation models on fisheye camera images. By utilizing the self-distillation technique, the AbsRel and SILog error metrics were reduced by 1.81% and 1.55% on the JBNU-Depth360 dataset. The experimental results demonstrated that the self-distillation technique is beneficial to improve the performance of depth-estimation models. [ABSTRACT FROM AUTHOR]

Monocular depth prediction research is essential for expanding meaning from 2D to 3D. Recent studies have focused on the application of a newly proposed encoder; however, the development within the self-supervised learning framework remains unexplored, an aspect critical for advancing foundational models of 3D semantic interpretation. Addressing the dynamic nature of encoder-based research, especially in performance evaluations for feature extraction and pre-trained models, this research proposes the switchable encoder learning framework (SELF). SELF enhances versatility by enabling the seamless integration of diverse encoders in a self-supervised learning context for depth prediction. This integration is realized through the direct transfer of feature information from the encoder and by standardizing the input structure of the decoder to accommodate various encoder architectures. Furthermore, the framework is extended and incorporated into an adaptable decoder for depth prediction and camera pose learning, employing standard loss functions. Comparative experiments with previous frameworks using the same encoder reveal that SELF achieves a 7% reduction in parameters while enhancing performance. Remarkably, substituting newly proposed algorithms in place of an encoder improves the outcomes as well as significantly decreases the number of parameters by 23%. The experimental findings highlight the ability of SELF to broaden depth factors, such as depth consistency. This framework facilitates the objective selection of algorithms as a backbone for extended research in monocular depth prediction. [ABSTRACT FROM AUTHOR]

We present an analysis of a self-supervised learning approach for monocular depth and ego-motion estimation. This is an important problem for computer vision systems of robots, autonomous vehicles and other intelligent agents, equipped only with monocular camera sensor. We have explored a number of neural network architectures that perform single-frame depth and multi-frame camera pose predictions to minimize photometric error between consecutive frames on a sequence of camera images. Unlike other existing works, our proposed approach called ERF-SfMLearner examines the influence of the deep neural network receptive field on the performance of depth and ego-motion estimation. To do this, we study the modification of network layers with two convolution operators with extended receptive field: dilated and deformable convolutions. We demonstrate on the KITTI dataset that increasing the receptive field leads to better metrics and lower errors both in terms of depth and ego-motion estimation. Code is publicly available at github.com/linukc/ERF-SfMLearner. [ABSTRACT FROM AUTHOR]

Obtaining dense depth ground-truth is not trivial, which leads to the introduction of self-supervised monocular depth estimation. Most self-supervised methods utilize the photometric loss as the primary supervisory signal to optimize a depth network. However, such self-supervised training often falls into an undesirable local minimum due to the ambiguity of the photometric loss. In this paper, we propose a novel self-distillation training scheme that provides a new self-supervision signal, depth consistency among different input resolutions, to the depth network. We further introduce a gradient masking strategy that adjusts the self-supervision signal of the depth consistency during back-propagation to boost the effectiveness of our depth consistency. Experiments demonstrate that our method brings meaningful performance improvements when applied to various depth network architectures. Furthermore, our method outperforms the existing self-supervised methods on KITTI, Cityscapes, and DrivingStereo datasets by a noteworthy margin. • A simple yet effective distillation is proposed for self-supervised monocular depth. • Resolving the resolution bias eventually improves depths in a target resolution. • The proposed depth consistency learning produces reliable self-supervision. • The gradients masking boosts the effectiveness of the depth consistency learning. • The depth consistency improves the generality of self-supervised depth networks. [ABSTRACT FROM AUTHOR]

Depth estimation is one of the crucial tasks for autonomous systems, which provides important information about the distance between the system and its surroundings. Traditionally, Light Detection and Ranging and stereo cameras have been used for distance measurement, despite the significant cost. In contrast, monocular cameras offer a more cost-effective solution, but lack inherent depth information. The synergy of big data and deep learning has led to various advanced architectures for monocular depth estimation. However, due to the characteristics of the monocular depth estimation case that is ill posed problem, we incorporate Attention Gates (AG) within an encoder-decoder based architecture. This helps prevent pattern recognition failures caused by variations in object sizes that share identical depth values. Our research involves evaluating popular pretrained architectures, assessing the impact of using AG, and creating effective head blocks to tackle depth estimation challenges. Notably, our approach demonstrates improved evaluation metrics on the DIODE dataset, positioning Attention U-Net as a promising solution. Therefore, utilizing the superior performance obtained by Attention U-Net in performing monocular depth estimation on low-cost autonomous systems could relatively reduce the cost of using lidar or stereo cameras in measuring distance.11 https://github.com/KrisnaPinasthika/Deciphering-Pixel-Insights

In recent years, there has been extensive research and application of unsupervised monocular depth estimation methods for intelligent vehicles. However, a major limitation of most existing approaches is their inability to predict absolute depth values in physical units, as they generally suffer from the scale problem. Furthermore, most research efforts have focused on ground vehicles, neglecting the potential application of these methods to unmanned aerial vehicles (UAVs). To address these gaps, this paper proposes a novel absolute depth estimation method specifically designed for flight scenes using a monocular vision sensor, in which a geometry-based scale recovery algorithm serves as a post-processing stage of relative depth estimation results with scale consistency. By exploiting the feature correspondence between successive images and using the pose data provided by equipped navigation sensors, the scale factor between relative and absolute scales is calculated according to a multi-view geometry model, and then absolute depth maps are generated by pixel-wise multiplication of relative depth maps with the scale factor. As a result, the unsupervised monocular depth estimation technology is extended from relative depth estimation in semi-structured scenes to absolute depth estimation in unstructured scenes. Experiments on the publicly available Mid-Air dataset and customized data demonstrate the effectiveness of our method in different cases and settings, as well as its robustness to navigation sensor noise. The proposed method only requires UAVs to be equipped with monocular camera and common navigation sensors, and the obtained absolute depth information can be directly used for downstream tasks, which is significant for this kind of vehicle that has rarely been explored in previous depth estimation studies.