Despite the great success of deep learning in stereo matching, recovering accurate and clearly-contoured disparity map is still challenging. Currently, L1 loss and cross-entropy loss are the two most widely used loss functions for training the stereo matching networks. Comparing with the former, the latter can usually achieve better results thanks to its direct constraint to the the cost volume. However, how to generate reasonable ground-truth distribution for this loss function remains largely under exploited. Existing works assume uni-modal distributions around the ground-truth for all of the pixels, which ignores the fact that the edge pixels may have multi-modal distributions. In this paper, we first experimentally exhibit the importance of correct edge supervision to the overall disparity accuracy. Then a novel adaptive multi-modal cross-entropy loss which encourages the network to generate different distribution patterns for edge and non-edge pixels is proposed. We further optimize the disparity estimator in the inference stage to alleviate the bleeding and misalignment artifacts at the edge. Our method is generic and can help classic stereo matching models regain competitive performance. GANet trained by our loss ranks 1st on the KITTI 2015 and 2012 benchmarks and outperforms state-of-the-art methods by a large margin. Meanwhile, our method also exhibits superior cross-domain generalization ability and outperforms existing generalization-specialized methods on four popular real-world datasets.
As a fundamental task for intelligent robots, visual SLAM has made great progress over the past decades. However, robust SLAM under highly weak-textured environments still remains very challenging. In this paper, we propose a novel visual SLAM system named RWT-SLAM to tackle this problem. We modify LoFTR network which is able to produce dense point matching under low-textured scenes to generate feature descriptors. To integrate the new features into the popular ORB-SLAM framework, we develop feature masks to filter out the unreliable features and employ KNN strategy to strengthen the matching robustness. We also retrained visual vocabulary upon new descriptors for efficient loop closing. The resulting RWT-SLAM is tested in various public datasets such as TUM and OpenLORIS, as well as our own data. The results shows very promising performance under highly weak-textured environments.
Visual relocalization aims to estimate the pose of a camera from one or more images. In recent years deep learning based pose regression methods have attracted many attentions. They feature predicting the absolute poses without relying on any prior built maps or stored images, making the relocalization very efficient. However, robust relocalization under environments with complex appearance changes and real dynamics remains very challenging. In this paper, we propose to enhance the distinctiveness of the image features by extracting the deep relationship among objects. In particular, we extract objects in the image and construct a deep object relation graph (ORG) to incorporate the semantic connections and relative spatial clues of the objects. We integrate our ORG module into several popular pose regression models. Extensive experiments on various public indoor and outdoor datasets demonstrate that our method improves the performance significantly and outperforms the previous approaches.
Monocular 3D object detection is an essential task in autonomous driving. However, most current methods consider each 3D object in the scene as an independent training sample, while ignoring their inherent geometric relations, thus inevitably resulting in a lack of leveraging spatial constraints. In this paper, we propose a novel method that takes all the objects into consideration and explores their mutual relationships to help better estimate the 3D boxes. Moreover, since 2D detection is more reliable currently, we also investigate how to use the detected 2D boxes as guidance to globally constrain the optimization of the corresponding predicted 3D boxes. To this end, a differentiable loss function, termed as Homography Loss, is proposed to achieve the goal, which exploits both 2D and 3D information, aiming at balancing the positional relationships between different objects by global constraints, so as to obtain more accurately predicted 3D boxes. Thanks to the concise design, our loss function is universal and can be plugged into any mature monocular 3D detector, while significantly boosting the performance over their baseline. Experiments demonstrate that our method yields the best performance (Nov. 2021) compared with the other state-of-the-arts by a large margin on KITTI 3D datasets.
In recent years 3D object detection from LiDAR point clouds has made great progress thanks to the development of deep learning technologies. Although voxel or point based methods are popular in 3D object detection, they usually involve time-consuming operations such as 3D convolutions on voxels or ball query among points, making the resulting network inappropriate for time critical applications. On the other hand, 2D view-based methods feature high computing efficiency while usually obtaining inferior performance than the voxel or point based methods. In this work, we present a real-time view-based single stage 3D object detector, namely CVFNet to fulfill this task. To strengthen the cross-view feature learning under the condition of demanding efficiency, our framework extracts the features of different views and fuses them in an efficient progressive way. We first propose a novel Point-Range feature fusion module that deeply integrates point and range view features in multiple stages. Then, a special Slice Pillar is designed to well maintain the 3D geometry when transforming the obtained deep point-view features into bird's eye view. To better balance the ratio of samples, a sparse pillar detection head is presented to focus the detection on the nonempty grids. We conduct experiments on the popular KITTI and NuScenes benchmark, and state-of-the-art performances are achieved in terms of both accuracy and speed.
Depth Completion can produce a dense depth map from a sparse input and provide a more complete 3D description of the environment. Despite great progress made in depth completion, the sparsity of the input and low density of the ground truth still make this problem challenging. In this work, we propose DenseLiDAR, a novel real-time pseudo-depth guided depth completion neural network. We exploit dense pseudo-depth map obtained from simple morphological operations to guide the network in three aspects: (1) Constructing a residual structure for the output; (2) Rectifying the sparse input data; (3) Providing dense structural loss for training the network. Thanks to these novel designs, higher performance of the output could be achieved. In addition, two new metrics for better evaluating the quality of the predicted depth map are also presented. Extensive experiments on KITTI depth completion benchmark suggest that our model is able to achieve the state-of-the-art performance at the highest frame rate of 50Hz. The predicted dense depth is further evaluated by several downstream robotic perception or positioning tasks. For the task of 3D object detection, 3~5 percent performance gains on small objects categories are achieved on KITTI 3D object detection dataset. For RGB-D SLAM, higher accuracy on vehicle's trajectory is also obtained in KITTI Odometry dataset. These promising results not only verify the high quality of our depth prediction, but also demonstrate the potential of improving the related downstream tasks by using depth completion results.
Although instance segmentation has made considerable advancement over recent years, it's still a challenge to design high accuracy algorithms with real-time performance. In this paper, we propose a real-time instance segmentation framework termed OrienMask. Upon the one-stage object detector YOLOv3, a mask head is added to predict some discriminative orientation maps, which are explicitly defined as spatial offset vectors for both foreground and background pixels. Thanks to the discrimination ability of orientation maps, masks can be recovered without the need for extra foreground segmentation. All instances that match with the same anchor size share a common orientation map. This special sharing strategy reduces the amortized memory utilization for mask predictions but without loss of mask granularity. Given the surviving box predictions after NMS, instance masks can be concurrently constructed from the corresponding orientation maps with low complexity. Owing to the concise design for mask representation and its effective integration with the anchor-based object detector, our method is qualified under real-time conditions while maintaining competitive accuracy. Experiments on COCO benchmark show that OrienMask achieves 34.8 mask AP at the speed of 42.7 fps evaluated with a single RTX 2080 Ti. The code is available at https://github.com/duwt/OrienMask.