3D Object Pose Estimation Research Articles

6D Object Pose Estimation Based on Cross-Modality Feature Fusion.

The 6D pose estimation using RGBD images plays a pivotal role in robotics applications. At present, after obtaining the RGB and depth modality information, most methods directly concatenate them without considering information interactions. This leads to the low accuracy of 6D pose estimation in occlusion and illumination changes. To solve this problem, we propose a new method to fuse RGB and depth modality features. Our method effectively uses individual information contained within each RGBD image modality and fully integrates cross-modality interactive information. Specifically, we transform depth images into point clouds, applying the PointNet++ network to extract point cloud features; RGB image features are extracted by CNNs and attention mechanisms are added to obtain context information within the single modality; then, we propose a cross-modality feature fusion module (CFFM) to obtain the cross-modality information, and introduce a feature contribution weight training module (CWTM) to allocate the different contributions of the two modalities to the target task. Finally, the result of 6D object pose estimation is obtained by the final cross-modality fusion feature. By enabling information interactions within and between modalities, the integration of the two modalities is maximized. Furthermore, considering the contribution of each modality enhances the overall robustness of the model. Our experiments indicate that the accuracy rate of our method on the LineMOD dataset can reach 96.9%, on average, using the ADD (-S) metric, while on the YCB-Video dataset, it can reach 94.7% using the ADD-S AUC metric and 96.5% using the ADD-S score (<2 cm) metric.

Self-supervised Vision Transformers for 3D pose estimation of novel objects

Object pose estimation is important for object manipulation and scene understanding. In order to improve the general applicability of pose estimators, recent research focuses on providing estimates for novel objects, that is, objects unseen during training. Such works use deep template matching strategies to retrieve the closest template connected to a query image, which implicitly provides object class and pose. Despite the recent success and improvements of Vision Transformers over CNNs for many vision tasks, the state of the art uses CNN-based approaches for novel object pose estimation. This work evaluates and demonstrates the differences between self-supervised CNNs and Vision Transformers for deep template matching. In detail, both types of approaches are trained using contrastive learning to match training images against rendered templates of isolated objects. At test time such templates are matched against query images of known and novel objects under challenging settings, such as clutter, occlusion and object symmetries, using masked cosine similarity. The presented results not only demonstrate that Vision Transformers improve matching accuracy over CNNs but also that for some cases pre-trained Vision Transformers do not need fine-tuning to achieve the improvement. Furthermore, we highlight the differences in optimization and network architecture when comparing these two types of networks for deep template matching.

OHO: A Multi-Modal, Multi-Purpose Dataset for Human-Robot Object Hand-Over.

In the context of collaborative robotics, handing over hand-held objects to a robot is a safety-critical task. Therefore, a robust distinction between human hands and presented objects in image data is essential to avoid contact with robotic grippers. To be able to develop machine learning methods for solving this problem, we created the OHO (Object Hand-Over) dataset of tools and other everyday objects being held by human hands. Our dataset consists of color, depth, and thermal images with the addition of pose and shape information about the objects in a real-world scenario. Although the focus of this paper is on instance segmentation, our dataset also enables training for different tasks such as 3D pose estimation or shape estimation of objects. For the instance segmentation task, we present a pipeline for automated label generation in point clouds, as well as image data. Through baseline experiments, we show that these labels are suitable for training an instance segmentation to distinguish hands from objects on a per-pixel basis. Moreover, we present qualitative results for applying our trained model in a real-world application.

Learning geometric consistency and discrepancy for category-level 6D object pose estimation from point clouds

Category-level 6D object pose estimation aims to predict the position and orientation of unseen object instances, which is a fundamental problem in robotic applications. Previous works mainly focused on exploiting visual cues from RGB images, while depth images received less attention. However, depth images contain rich geometric attributes about the object’s shape, which are crucial for inferring the object’s pose. This work achieves category-level 6D object pose estimation by performing sufficient geometric learning from depth images represented by point clouds. Specifically, we present a novel geometric consistency and geometric discrepancy learning framework called CD-Pose to resolve the intra-category variation, inter-category similarity, and objects with complex structures. Our network consists of a Pose-Consistent Module and a Pose-Discrepant Module. First, a simple MLP-based Pose-Consistent Module is utilized to extract geometrically consistent pose features of objects from the pre-computed object shape priors for each category. Then, the Pose-Discrepant Module, designed as a multi-scale region-guided transformer network, is dedicated to exploring each instance’s geometrically discrepant features. Next, the NOCS model of the object is reconstructed according to the integration of consistent and discrepant geometric representations. Finally, 6D object poses are obtained by solving the similarity transformation between the reconstruction and the observed point cloud. Experiments on the benchmark datasets show that our CD-Pose produces superior results to state-of-the-art competitors.

OTOT: An online training and offline testing system for 6D Object Pose Estimation

OTOT: An online training and offline testing system for 6D Object Pose Estimation

YOLOPose V2: Understanding and improving transformer-based 6D pose estimation

6D object pose estimation is a crucial prerequisite for autonomous robot manipulation applications. The state-of-the-art models for pose estimation are convolutional neural network (CNN)-based. Lately, Transformers, an architecture originally proposed for natural language processing, is achieving state-of-the-art results in many computer vision tasks as well. Equipped with the multi-head self-attention mechanism, Transformers enable simple single-stage end-to-end architectures for learning object detection and 6D object pose estimation jointly. In this work, we propose YOLOPose (short form for You Only Look Once Pose estimation), a Transformer-based multi-object 6D pose estimation method based on keypoint regression and an improved variant of the YOLOPose model. In contrast to the standard heatmaps for predicting keypoints in an image, we directly regress the keypoints. Additionally, we employ a learnable orientation estimation module to predict the orientation from the keypoints. Along with a separate translation estimation module, our model is end-to-end differentiable. Our method is suitable for real-time applications and achieves results comparable to state-of-the-art methods. We analyze the role of object queries in our architecture and reveal that the object queries specialize in detecting objects in specific image regions. Furthermore, we quantify the accuracy trade-off of using datasets of smaller sizes to train our model.

Recent Advances and Perspectives in Deep Learning Techniques for 3D Point Cloud Data Processing

In recent years, deep learning techniques for processing 3D point cloud data have seen significant advancements, given their unique ability to extract relevant features and handle unstructured data. These techniques find wide-ranging applications in fields like robotics, autonomous vehicles, and various other computer-vision applications. This paper reviews the recent literature on key tasks, including 3D object classification, tracking, pose estimation, segmentation, and point cloud completion. The review discusses the historical development of these methods, explores different model architectures, learning algorithms, and training datasets, and provides a comprehensive summary of the state-of-the-art in this domain. The paper presents a critical evaluation of the current limitations and challenges in the field, and identifies potential areas for future research. Furthermore, the emergence of transformative methodologies like PoinTr and SnowflakeNet is examined, highlighting their contributions and potential impact on the field. The potential cross-disciplinary applications of these techniques are also discussed, underscoring the broad scope and impact of these developments. This review fills a knowledge gap by offering a focused and comprehensive synthesis of recent research on deep learning techniques for 3D point cloud data processing, thereby serving as a useful resource for both novice and experienced researchers in the field.

MORE: simultaneous multi-view 3D object recognition and pose estimation

Simultaneous object recognition and pose estimation are two key functionalities for robots to safely interact with humans as well as environments. Although both object recognition and pose estimation use visual input, most state of the art tackles them as two separate problems since the former needs a view-invariant representation, while object pose estimation necessitates a view-dependent description. Nowadays, multi-view convolutional neural network (MVCNN) approaches show state-of-the-art classification performance. Although MVCNN object recognition has been widely explored, there has been very little research on multi-view object pose estimation methods, and even less on addressing these two problems simultaneously. The pose of virtual cameras in MVCNN methods is often pre-defined in advance, leading to bound the application of such approaches. In this paper, we propose an approach capable of handling object recognition and pose estimation simultaneously. In particular, we develop a deep object-agnostic entropy estimation model, capable of predicting the best viewpoints of a given 3D object. The obtained views of the object are then fed to the network to simultaneously predict the pose and category label of the target object. Experimental results showed that the views obtained from such positions are descriptive enough to achieve a good accuracy score. Furthermore, we designed a real-life serve drink scenario to demonstrate how well the proposed approach worked in real robot tasks. Code is available online at: https://github.com/SubhadityaMukherjee/more_mvcnn.

SANet: A novel segmented attention mechanism and multi-level information fusion network for 6D object pose estimation

Reliably and rapidly estimating the 6D position of an object is a critical challenge when using Internet of Things (IoT) technologies for monitoring. Nowadays, the prevalent 6D pose estimation architecture is based on a two-stage technique, which requires a significant amount of time for both training and deploying the algorithm in actual applications. Additionally, the majority of approaches include intricate high-low level features in the network that greatly influence training but contribute less to testing. To enable more accurate 6D object pose estimation while shortening the deployment time, we used a single-stage end-to-end algorithm to design the network. In this paper, we propose SANet, which is composed of a segmented attention module and a multi-level information fusion module. Specifically, by extracting high-level semantic information from images before fusing them to the decoder, and by removing redundant information using the multi-level information fusion module, the feature fusion complexity of the model is reduced by extracting high level features. In addition, the segmented attention module can suppress unreliable information to enhance network learning of channel and spatial information, enabling the network to more accurately understand the geometric aspects of the object. Extensive experiments on LM and LMO datasets demonstrate that our method outperforms state-of-the-art baselines, ranking 1st in both speed and accuracy.

Imitrob: Imitation Learning Dataset for Training and Evaluating 6D Object Pose Estimators

This paper introduces a dataset for training and evaluating methods for 6D pose estimation of hand-held tools in task demonstrations captured by a standard RGB camera. Despite the significant progress of 6D pose estimation methods, their performance is usually limited for heavily occluded objects, which is a common case in imitation learning, where the object is typically partially occluded by the manipulating hand. Currently, there is a lack of datasets that would enable the development of robust 6D pose estimation methods for these conditions. To overcome this problem, we collect a new dataset (Imitrob) aimed at 6D pose estimation in imitation learning and other applications where a human holds a tool and performs a task. The dataset contains image sequences of nine different tools and twelve manipulation tasks with two camera viewpoints, four human subjects, and left/right hand. Each image is accompanied by an accurate ground truth measurement of the 6D object pose obtained by the HTC Vive motion tracking device. The use of the dataset is demonstrated by training and evaluating a recent 6D object pose estimation method (DOPE) in various setups.

Self-Supervised Category-Level 6D Object Pose Estimation With Optical Flow Consistency

Category-level 6D object pose estimation aims at determining the pose of an object of a given category. Most current state-of-the-art methods require a significant amount of real training data to supervise their models. Moreover, annotating the 6D pose is very time consuming, error-prone, and it does not scale well to a large amount of object classes. Therefore, a handful of methods have recently been proposed to use unlabelled data to establish weak supervision. In this letter we propose a self-supervised method that leverages the 2D optical flow as a proxy for supervising the 6D pose. To this purpose, we estimate the 2D optical flow between consecutive frames based on the pose estimation. Then, we harness an off-the-shelf optical flow method to enable weak supervision using a 2D-3D optical flow based consistency loss. Experiments show that our approach for self-supervised learning yields state-of-the-art performance on the NOCS benchmark, and it reaches comparable results with some fully-supervised approaches.

6D Object Localization in Car-Assembly Industrial Environment

In this work, a visual object detection and localization workflow integrated into a robotic platform is presented for the 6D pose estimation of objects with challenging characteristics in terms of weak texture, surface properties and symmetries. The workflow is used as part of a module for object pose estimation deployed to a mobile robotic platform that exploits the Robot Operating System (ROS) as middleware. The objects of interest aim to support robot grasping in the context of human-robot collaboration during car door assembly in industrial manufacturing environments. In addition to the special object properties, these environments are inherently characterised by cluttered background and unfavorable illumination conditions. For the purpose of this specific application, two different datasets were collected and annotated for training a learning-based method that extracts the object pose from a single frame. The first dataset was acquired in controlled laboratory conditions and the second in the actual indoor industrial environment. Different models were trained based on the individual datasets and a combination of them were further evaluated in a number of test sequences from the actual industrial environment. The qualitative and quantitative results demonstrate the potential of the presented method in relevant industrial applications.

Deep learning for 6D pose estimation of objects — A case study for autonomous driving

Nowadays, the potential benefits and implementation of autonomous driving have attracted widespread attention from both industry and academia. This study will solve view-invariant object detection and semantic key-point pose assumptions from a single RGB image. A machine learning method for estimating the absolute pose of an on-road vehicle for autonomous driving from monocular vision alone without the help of additional sensors is a complex task. The main purpose of this work is to identify other vehicles on the road and estimate their exact angular position from a single image with improved accuracy. The focus of the study is to create a new algorithm by applying a potentially deep convoluted neural network followed by a repetitive neural structure for more accurate 6D pose inference. A 6D pose hypothesis is presented in this study, based on a deep hybrid architecture for individual vehicles of an end-to-end approach to a task consisting of a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN). In this work, we will use a large-scale dataset consistent with the understanding of a 3D car instance called ApolloCar3D. The data set contains 5,277 real-life street scenes with examples of about 60K cars. By comparison, the ApolloCar3D is twenty times larger than the PASCAL3D+ and KITTI datasets. Ultimately, the idea is to efficiently eliminate motionless cars and predict the next pose given in the speed context, allowing a comprehensive evaluation, and passing the output through LSTM (long short-term memory) with an additional filter layer. The new filter added to the LSTM will efficiently filter and isolate stationary or parking vehicles and focus on on-road vehicles. Since the LSTM has a non-linear high-dimensional hidden memory state, it can preserve the past continuity of each generation’s data history and pay more attention to those road vehicles rather than parked or stationary vehicles to act accordingly. So for each new vehicle, the pose estimator classifier can use LSTM memory and compare the historical pose with the newly filtered data. The successful implementation of this innovative concept will lead to significant improvements in the real-life traffic situation in the field of computer vision and autonomous driving.

Weak6D: Weakly Supervised 6D Pose Estimation With Iterative Annotation Resolver

6D object pose estimation is an essential task in vision-based robotic grasping and manipulation. Prior works always train models with a large number of pose annotated images, limiting the efficiency of model transfer between different scenarios. This letter presents an end-to-end model named <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Weak6D</i> , which could be learned with unannotated RGB-D data. The core of the proposed approach is the novel optimizing method Iterative Annotation Resolver, which has the ability to directly utilize the captured RGB-D data through the training process. Furthermore, we employ a weak refinement loss to optimize the pose estimation network with refined object poses. We evaluated the proposed <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Weak6D</i> in the YCB-Video dataset, and experimental results show our model achieved practical results without annotated data.

Object-Aware 3D Scene Reconstruction from Single 2D Images of Indoor Scenes

Recent studies have shown that deep learning achieves excellent performance in reconstructing 3D scenes from multiview images or videos. However, these reconstructions do not provide the identities of objects, and object identification is necessary for a scene to be functional in virtual reality or interactive applications. The objects in a scene reconstructed as one mesh are treated as a single object, rather than individual entities that can be interacted with or manipulated. Reconstructing an object-aware 3D scene from a single 2D image is challenging because the image conversion process from a 3D scene to a 2D image is irreversible, and the projection from 3D to 2D reduces a dimension. To alleviate the effects of dimension reduction, we proposed a module to generate depth features that can aid the 3D pose estimation of objects. Additionally, we developed a novel approach to mesh reconstruction that combines two decoders that estimate 3D shapes with different shape representations. By leveraging the principles of multitask learning, our approach demonstrated superior performance in generating complete meshes compared to methods relying solely on implicit representation-based mesh reconstruction networks (e.g., local deep implicit functions), as well as producing more accurate shapes compared to previous approaches for mesh reconstruction from single images (e.g., topology modification networks). The proposed method was evaluated on real-world datasets. The results showed that it could effectively improve the object-aware 3D scene reconstruction performance over existing methods.

Geometric-aware dense matching network for 6D pose estimation of objects from RGB-D images

6D pose estimation for certain targets from RGB-D images is a fundamental problem in computer vision. Current methods emphasize learning the overall expression of the targets, which leads to poor performance under occlusion and truncation conditions. In this paper, we propose using a geometric-aware dense matching network to obtain visible dense correspondences between a RGB-D image and 3D model to address difficult predictions from unseen keypoints. Two geometrical structures are considered for dense matching. (1) The neighbor area of the correspondences is treated as suboptimal matches in addition to the correspondence to reduce the influence of the error caused by ground truth calibration. (2) The distance consistency of the correspondences is leveraged to eliminate the ambiguity from the symmetrical objects. Experiments on LM-O dataset (77.1% ADD(S)-0.1d) and YCB-V dataset (97.6% ADD(S)) show the effectiveness and advantages of our proposed method.11The source code will soon be available at https://github.com/Ray0089/geometric-aware-dense-matching.

Ambiguity-Aware Multi-Object Pose Optimization for Visually-Assisted Robot Manipulation

6D object pose estimation aims to infer the relative pose between the object and the camera using a single image or multiple images. Most works have focused on predicting the object pose without associated uncertainty under occlusion and structural ambiguity (symmetricity). However, these works demand prior information about shape attributes, and this condition is hardly satisfied in reality; even asymmetric objects may be symmetric under the viewpoint change. In addition, acquiring and fusing diverse sensor data is challenging when extending them to robotics applications. Tackling these limitations, we present an ambiguity-aware 6D object pose estimation network, PrimA6D++, as a generic uncertainty prediction method. The major challenges in pose estimation, such as occlusion and symmetry, can be handled in a generic manner based on the measured ambiguity of the prediction. Specifically, we devise a network to reconstruct the three rotation axis primitive images of a target object and predict the underlying uncertainty along each primitive axis. Leveraging the estimated uncertainty, we then optimize multi-object poses using visual measurements and camera poses by treating it as an object SLAM problem. The proposed method shows a significant performance improvement in T-LESS and YCB-Video datasets. We further demonstrate real-time scene recognition capability for visually-assisted robot manipulation.

Towards Robot-Assisted Data Generation with Minimal User Interaction for Autonomously Training 6D Pose Estimation in Operational Environments

Recently, deep neural networks achieved state-of-the-art results in the subject of 6D object pose estimation for robot manipulation. However, those deep learning methods require expensive training data. Current methods for cost reduction use synthetic data, but rely on expert knowledge and suffer from the domain gap when shifting to the real application. Here, we present a proof of concept for a novel approach towards generating annotated training data for 6D object pose estimation. This approach is designed for learning new objects in operational environments while requiring little interaction and no expertise on the part of the user.

VP-KLNet: efficient 6D object pose estimation with an enhanced vector-field prediction network and a keypoint localization network

6D object pose estimation is essential for many applications with high demands in accuracy and speed. Compared with end-to-end approaches, pixel-wise voting network (PVNet), a vector-field based two-stage approach, has shown the superiority in accuracy but the inferiority in speed because of the time-consuming RANSAC-based voting strategy. To resolve this problem, we propose an efficient deep architecture that consists of an enhanced vector-field prediction network (VPNet) and a keypoint localization network (KLNet) and call it VP-KLNet. Specifically, the KLNet replaces PVNet’s time-consuming voting scheme by directly regressing 2D keypoints from the vector fields, which significantly improves the inference speed. Furthermore, to capture multiscale contextual information, we embed the pyramid pooling module between the encoder and decoder in VPNet to obtain more accurate object segmentation and vector-field prediction, with negligible speed loss. Experiments demonstrate that our method has more than 50% improved in the running speed to the baseline method PVNet and achieves comparable accuracy with the state-of-the-art methods on the LINEMOD and occlusion LINEMOD datasets.

Category-Level 6D Object Pose Estimation With Structure Encoder and Reasoning Attention

Category-level 6D object pose estimation has gained popularity and it is still challenging due to the diversity of different instances within the same category. In this paper, a novel category-level 6D object pose estimation framework with structure encoder and reasoning attention is proposed. A structure autoencoder is introduced to mine the shared structure features in the color images within the same category, via a distinct learning strategy that recovers the image of another instance but with the most similar pose to the input. On this basis, a reasoning attention decoder and full connected layers are stacked to form a rotation prediction network, where the structure features and 3D shape features are integrated and projected to a semantic space. The semantic space includes observed patterns and learnable patterns, which are better learned by adding a shortcut connection branch parallel to reasoning attention decoder with gradient decouple. Further reasoning based on these patterns endows the decoder with powerful feature representation. Without 3D object models, the proposed method models the attributes of category implicitly in the semantic space and better performance of 6D object pose estimation is guaranteed by reasoning on this space. The effectiveness of the proposed method is verified by the results on public datasets and actual experiments.