Search

Image stitching is a classical and challenging technique in computer vision, which aims to generate an image with a wide field of view. The traditional methods heavily depend on feature detection and require the feature points to be dense and evenly distributed in the image, leading to poor robustness in low-texture scenes. Learning methods are rarely studied due to the unavailability of ground truth stitched results, showing unreliable performance on real-world datasets. In this paper, we propose an image stitching learning framework, which consists of a multi-scale deep homography module and an edge-preserved deformation module. First, we design a multi-scale deep homography module to estimate the accurate homography progressively from coarse to fine. After that, an edge-preserved deformation module is designed to learn the deformation rules of image stitching from edge to content, generating the stitched image with artifacts eliminated. Besides, the proposed supervised learning framework can stitch images of arbitrary resolutions and demonstrate good generalization capability in real-world images. Experiments show that our superiority to the existing homography solutions and image stitching algorithms. [ABSTRACT FROM AUTHOR]

Fisheye lenses are widely applied in the self-driving system due to its large field of view (FOV), which also induces the strong radial distortion in captured data. Distortion rectification is a crucial preprocessing step for scene perception since the non-linear projection heavily destroys the original geometric distribution of objects. Previous methods rectify the distorted image with a plausible global distribution, but they perform poorly on the local appearance such as the detailed texture of vehicles and pedestrians, which are hard to meet the fine requirements of object detection and semantic recognition. To solve this problem, we propose a joint distortion rectification (DR) and super-resolution (SR) end-to-end framework to generate the realistic rectification result with a visually pleasing local appearance. To the best of our knowledge, this is the first parametric framework that combines the challenging techniques DR and SR in the field of computer vision, which learns the accurate mapping function from the distorted and low-resolution domain into the corrected and high-resolution domain simultaneously. Moreover, the devised network architecture using the channel attention mechanism and object-aware loss function enables better-reconstructed scenes compared with the state-of-the-art methods. Experimental results show that our approach achieves superior rectification and enhancement performance in both quantitative measure and visual qualitative appearance, contributing a comprehensive and promising vision algorithm to the perception of the self-driving scene. [ABSTRACT FROM AUTHOR]

Recently, the pseudo-LiDAR gains increasing attention since it provides the probability of replacing the expensive LiDAR with the camera in autonomous driving. However, due to the fixed field of view of the camera, the pseudo-LiDAR point cloud suffers from the limited spatial range. In this paper, we present a novel pseudo-LiDAR point cloud magnification algorithm, aiming to extrapolate the narrow baseline and further bridge the gap between LiDAR and camera. To achieve this goal, we design a complete pipeline consisting of the hybrid view synthesis module, stereo depth estimation module, image-depth associated stitching module, and magnified pseudo-LiDAR point cloud transformation module. Considering the artifacts existed in generated views, we propose a region-aware restoration approach to obtain more realistic synthesized results. To the best of our knowledge, this is the first work for the pseudo-LiDAR point cloud magnification, which shows appealing and meaningful applications in 3D spatial perception systems only equipped with cameras. Experimental results on KITTI benchmark demonstrate that our algorithm can effectively magnify pseudo-LiDAR point cloud with a wider field of view, note that we only use two images captured by stereo cameras. [ABSTRACT FROM AUTHOR]

Image denoising which aims to restore a high-quality image from the noisy version is one of the most challenging tasks in the low-level computer vision tasks. In this paper, we propose a multi-stage progressive denoising network (MSPNet) and decompose the denoising task into some sub-tasks to progressively remove noise. Specifically, MSPNet is composed of three denoising stages. Each stage combines a feature extraction module (FEM) and a mutual-learning fusion module (MFM). In the feature extraction module, an encoder-decoder architecture is employed to learn non-local contextualized features, and the channel attention blocks (CAB) are utilized to retain the local information of the image. In the mutual-learning fusion module, the criss-cross attention is introduced to balance the image spatial details and the contextualized information. Compared with the state-of-the-art works, experimental results show that MSPNet achieves notable improvements on both objective and subjective evaluations. [ABSTRACT FROM AUTHOR]

The resolution of the depth map captured by a time-of-flight camera is rather limited; hence, it cannot be directly used to build a high-quality 3D model, together with an associated high-resolution texture image. In this paper, an improved depth up-sampling algorithm that considers two important characteristics of the depth map is proposed. First, the invariance of the fractal dimension between the low-resolution and high-resolution depth maps is maintained. Second, the boundaries of the depth map and its corresponding texture image are made to be consistent to provide a good synthesized virtual view. The experimental results show that the proposed method achieves superior performance for the up-sampled depth map and synthesized virtual view compared with classical approaches. [ABSTRACT FROM AUTHOR]