Rethinking Training Strategy in Stereo Matching

Triangle-Mesh-Rasterization-Projection (TMRP): An Algorithm to Project a Point Cloud onto a Consistent, Dense and Accurate 2D Raster Image

The projection of a point cloud onto a 2D camera image is relevant in the case of various image analysis and enhancement tasks, e.g., (i) in multimodal image processing for data fusion, (ii) in robotic applications and in scene analysis, and (iii) for deep neural networks to generate real datasets with ground truth. The challenges of the current single-shot projection methods, such as simple state-of-the-art projection, conventional, polygon, and deep learning-based upsampling methods or closed source SDK functions of low-cost depth cameras, have been identified. We developed a new way to project point clouds onto a dense, accurate 2D raster image, called Triangle-Mesh-Rasterization-Projection (TMRP). The only gaps that the 2D image still contains with our method are valid gaps that result from the physical limits of the capturing cameras. Dense accuracy is achieved by simultaneously using the 2D neighborhood information (rx,ry) of the 3D coordinates in addition to the points P(X,Y,V). In this way, a fast triangulation interpolation can be performed. The interpolation weights are determined using sub-triangles. Compared to single-shot methods, our algorithm is able to solve the following challenges. This means that: (1) no false gaps or false neighborhoods are generated, (2) the density is XYZ independent, and (3) ambiguities are eliminated. Our TMRP method is also open source, freely available on GitHub, and can be applied to almost any sensor or modality. We also demonstrate the usefulness of our method with four use cases by using the KITTI-2012 dataset or sensors with different modalities. Our goal is to improve recognition tasks and processing optimization in the perception of transparent objects for robotic manufacturing processes.

An application of stereo matching algorithm based on transfer learning on robots in multiple scenes

Robot vision technology based on binocular vision holds tremendous potential for development in various fields, including 3D scene reconstruction, target detection, and autonomous driving. However, current binocular vision methods used in robotics engineering have limitations such as high costs, complex algorithms, and low reliability of the generated disparity map in different scenes. To overcome these challenges, a cross-domain stereo matching algorithm for binocular vision based on transfer learning was proposed in this paper, named Cross-Domain Adaptation and Transfer Learning Network (Ct-Net), which has shown valuable results in multiple robot scenes. First, this paper introduces a General Feature Extractor to extract rich general feature information for domain adaptive stereo matching tasks. Then, a feature adapter is used to adapt the general features to the stereo matching network. Furthermore, a Domain Adaptive Cost Optimization Module is designed to optimize the matching cost. A disparity score prediction module was also embedded to adaptively adjust the search range of disparity and optimize the cost distribution. The overall framework was trained using a phased strategy, and ablation experiments were conducted to verify the effectiveness of the training strategy. Compared with the prototype PSMNet, on KITTI 2015 benchmark, the 3PE-fg of Ct-Net in all regions and non-occluded regions decreased by 19.3 and 21.1% respectively, meanwhile, on the Middlebury dataset, the proposed algorithm improves the sample error rate at least 28.4%, which is the Staircase sample. The quantitative and qualitative results obtained from Middlebury, Apollo, and other datasets demonstrate that Ct-Net significantly improves the cross-domain performance of stereo matching. Stereo matching experiments in real-world scenes have shown that it can effectively address visual tasks in multiple scenes.

Multiple Frequency Inputs and Context-Guided Attention Network for Stereo Disparity Estimation

Deep learning-based methods have made remarkable progress for stereo matching in terms of accuracy. However, two issues still hinder producing a perfect disparity map: (1) blurred boundaries and the discontinuous disparity of a continuous region on disparity estimation maps, and (2) a lack of effective means to restore resolution precisely. In this paper, we propose to utilize multiple frequency inputs and an attention mechanism to construct the deep stereo matching model. Specifically, high-frequency and low-frequency information of the input image together with the RGB image are fed into a feature extraction network with 2D convolutions. It is conducive to produce a distinct boundary and continuous disparity of the smooth region on disparity maps. To regularize the 4D cost volume for disparity regression, we propose a 3D context-guided attention module for stacked hourglass networks, where high-level cost volumes as context guide low-level features to obtain high-resolution yet precise feature maps. The proposed approach achieves competitive performance on SceneFlow and KITTI 2015 datasets.