Search

Hand pose estimation is vital for various applications, including virtual reality (VR), augmented reality (AR), gesture recognition, human-computer interaction (HCI), and robotics. Achieving accurate and real-time hand pose estimation is challenging due to factors such as the high degree of articulation in the human hand and the variability in hand shapes and sizes. While multimodal data offers advantages, developing a fast and resource-efficient hand pose estimation system remains challenging. Current state-of-the-art methods often require powerful graphics processing units (GPUs) for high performance, limiting deployment on edge platforms with limited computational resources. There is a critical need for higher efficiency without compromising accuracy, especially in real-world applications like mobile devices and embedded systems. Additionally, real-time performance is essential for practical applications, where systems must respond immediately to user interactions. Unfortunately, most current methods struggle to achieve real-time speeds, even on powerful GPUs, let alone on resource-constrained devices. To address these challenges, we propose an efficient hand pose estimation system that leverages both red-green-blue (RGB) and depth (RGBD) data through a unified fusion strategy. Our method combines appearance and geometric data early in the processing pipeline, significantly reducing computational complexity while maintaining real-time performance on resource-constrained devices. Experimental results show that the proposed model runs at over 110 fps on GPU, and 30 fps on the edge platform of NVidia Jetson NX Xavier, which is 4 to 5 times faster than existing methods, while achieving competitive accuracy.

Monocular Depth Estimation (MDE) and Object Pose Estimation (OPE) are important tasks in visual scene understanding. Traditionally, these challenges have been addressed independently, with separate deep neural networks designed for each task. However, we contend that MDE, which provides information about object distances from the camera, and OPE, which focuses on determining precise object position and orientation, are inherently connected. Combining these tasks in a unified approach facilitates the integration of spatial context, offering a more comprehensive understanding of object distribution in three-dimensional space. Consequently, this work addresses both challenges simultaneously, treating them as a multi-task learning problem. Our proposed solution is encapsulated in a Unified Framework for Depth-Assisted Monocular Object Pose Estimation. Leveraging Red-Green-Blue (RGB) images as input, our framework estimates pose of multiple object instances alongside an instance-level depth map. During training, we utilize both depth and color images, but during inference, the model relies exclusively on color images. To enhance the depth-aware features crucial for robust object pose estimation, we introduce a depth estimation branch supervised by depth images. These features undergo further refinement through a cross-task attention module, contributing to the innovation of our method in significantly improving feature discriminability and robustness in object pose estimation. Through extensive experiments, our approach demonstrates competitive performance compared to state-of-the-art methods in object pose estimation. Moreover, our method operates in real-time, underscoring its efficiency and practical applicability in various scenarios. This unified framework not only advances the state of the art in monocular depth estimation and object pose estimation but also underscores the potential of multi-task learning for enhancing the understanding of complex visual scenes.

The task of object pose estimation in computer vision heavily relies on both color (RGB) and depth (D) images to provide crucial appearance and geometric information, assisting algorithms in understanding occlusions and object geometry, thereby enhancing accuracy. However, the dependency on specialized sensors capable of capturing depth poses challenges in terms of cost and availability. Consequently, researchers are exploring methods to estimate object poses solely from RGB images. Nevertheless, this approach encounters difficulties in handling occlusions, discerning object geometry, and resolving ambiguities arising from similar color or texture patterns. This paper introduces a novel geometry-aware method for object pose estimation utilizing RGB images as input to determine the poses of multiple object instances. Our approach leverages both depth and color images during training but only relies on color images during inference. Departing from traditional depth sensors, our method computes predicted point clouds directly from estimated depth images derived from RGB inputs. A key innovation lies in the formulation of a multi-scale fusion module adept at seamlessly integrating features extracted from RGB images with those inferred from the predicted point clouds. This fusion process significantly fortifies the pose estimation pipeline by harnessing the strengths of both modalities, resulting in notably improved object poses. Extensive experimentation demonstrates that our approach markedly outperforms state-of-the-art RGB-based methods on Occluded-LINEMOD and YCB-Video datasets. Moreover, our method achieves competitive results compared to RGB-D approaches that necessitate both RGB and depth data from physical sensors.

Grasp detection plays a pivotal role in robotic manipulation, allowing robots to interact with and manipulate objects in their surroundings. Traditionally, this has relied on three-dimensional (3D) point cloud data acquired from specialized depth cameras. However, the limited availability of such sensors in real-world scenarios poses a significant challenge. In many practical applications, robots operate in diverse environments where obtaining high-quality 3D point cloud data may be impractical or impossible. This paper introduces an innovative approach to grasp generation using color images, thereby eliminating the need for dedicated depth sensors. Our method capitalizes on advanced deep learning techniques for depth estimation directly from color images. Instead of relying on conventional depth sensors, our approach computes predicted point clouds based on estimated depth images derived directly from Red-Green-Blue (RGB) input data. To our knowledge, this is the first study to explore the use of predicted depth data for grasp detection, moving away from the traditional dependence on depth sensors. The novelty of this work is the development of a fusion module that seamlessly integrates features extracted from RGB images with those inferred from the predicted point clouds. Additionally, we adapt a voting mechanism from our previous work (VoteGrasp) to enhance robustness to occlusion and generate collision-free grasps. Experimental evaluations conducted on standard datasets validate the effectiveness of our approach, demonstrating its superior performance in generating grasp configurations compared to existing methods. With our proposed method, we achieved a significant 4% improvement in average precision compared to state-of-the-art grasp detection methods. Furthermore, our method demonstrates promising practical viability through real robot grasping experiments, achieving an impressive 84% success rate.

Hand-object configuration recovery is an important task in computer vision. The estimation of pose and shape for both hands and objects during interactive scenarios has various applications, particularly in augmented reality, virtual reality, or imitation-based robot learning. The problem is particularly challenging when the hand is interacting with objects in the environment, as this setting features both extreme occlusions and non-trivial shape deformations. While existing works treat the problem of estimating hand configurations (that is pose and shape parameters) in isolation from the recovery of parameters related to the object acted upon, we stipulate that the two problems are related and can be solved more accurately concurrently. We introduce an approach that jointly learns the features of hand and object from color and depth (RGB-D) images. Our approach fuses appearance and geometric features in an adaptive manner which allows us to accent or suppress features that are more meaningful for the upstream task of hand-object configuration recovery. We combine a deep Hough voting strategy that builds on our adaptive features with a graph convolutional network (GCN) to learn the interaction relationships between the hand and held object shapes during interaction. Experimental results demonstrate that our proposed approach consistently outperforms state-of-the-art methods on popular datasets.

Accurate detection and estimation of pallet poses from color and depth data (RGB-D) are integral components many in advanced warehouse intelligent systems. State-of-the art object pose estimation methods follow a two-stage process, relying on off-the-shelf segmentation or object detection in the initial stage and subsequently predicting the pose of objects using cropped images. The cropped patches may include both the target object and irrelevant information, such as background or other objects, leading to challenges in handling pallets in warehouse settings with heavy occlusions from loaded objects. In this study, we propose an innovative deep learning-based approach to address the occlusion problem in pallet pose estimation from RGB-D images. Inspired by the selective attention mechanism in human perception, our developed model learns to identify and attenuate the significance of features in occluded regions, focusing on the visible and informative areas for accurate pose estimation. Instead of directly estimating pallet poses from cropped patches as in existing methods, we introduce two feature map re-weighting modules with cross-modal attention. These modules effectively filter out features from occluded regions and background, enhancing pose estimation accuracy. Furthermore, we introduce a large-scale annotated pallet dataset specifically designed to capture occlusion scenarios in warehouse environments, facilitating comprehensive training and evaluation. Experimental results on the newly collected pallet dataset show that our proposed method increases accuracy by 13.5% compared to state-of-the-art methods.