Fig 1.
The disparity estimation results of the PSMNet algorithm in the ill-conditioned region.
The error map scales linearly between 0 (black) and 5 (white) pixel errors. The green box in the figure shows a large error in the disparity estimation of the vehicle and wall reflection areas.
Table 1.
Parameters of the improved PSMNet architecture.
Fig 2.
Schematic diagram of the atrous convolution.
Fig 3.
The pipeline of the proposed improved PSMNet network.
The left and right images are input into two weighted residual convolution neural networks used to extract features, and the ASPP module is used to obtain the contextual information of images. Then, the left and right image features are connected to form a 4-D cost space, and the costs are regularized by a multiscale 3D CNN network. Finally, an accurate disparity map is obtained by disparity regression.
Table 2.
The experiment of ASPP structure on Scene Flow.
Table 3.
The results of the ablation comparison on Scene Flow.
Fig 4.
Qualitative evaluation results on Scene Flow.
The first column shows the left images, the second column shows the ground truths, the third column shows the disparity maps estimated using PSMNet, and the fourth column shows the results of our model.
Fig 5.
Qualitative evaluation results on the KITTI 2012 dataset.
The first column shows the left images. For each photo, the first row shows the pseudocolor images for predicting the disparity map, and the second row shows the error maps. From left to right are the results of GC-NET [18], PSMNet [13], and our method, respectively.
Table 4.
Comparison of the EPEs of different networks on the Scene Flow dataset.
Fig 6.
Qualitative evaluation results on the KITTI 2015 dataset.
The first column shows the left images. For each photo, the first row shows the pseudocolor images for predicting the disparity map, and the second row shows the error maps. From left to right are the results of GC-NET [18], PSMNet [13], and our method, respectively.
Table 5.
KITTI 2012 quantitative assessment results.
Table 6.
KITTI 2015 quantitative assessment results.