Basic understanding of stereo depth estimation (Deep Learning)
I've summarized the flow of Depth (Disparity) estimation by Stereo Camera.
See the previous article for general stereo matching https://qiita.com/minh33/items/55717aa1ace9d7f7e7dd
Depth Estimating process 1. Feature Extraction => Extract features using the same network and the same weight on the left and right (WxHxC). ex) intensity, shape, ..... etc 2. Cost Volume => Create Disparity Channel by shifting Pixel (DxWxHxC). D is an arbitrary value, and you decide the maximum Disparity (pixel) that you can take. If it is too large, it will be heavy, and if it is too small, close objects will not be matched. 3. 3D Feature Matching => Learn to get a large value where the right and left features are close by convolving (DxWxHxC) with the obtained features (DxWxHx1) 4. Disparity Regression => Convert from (DxWxH) to (1xWxH) to find the final Disparity 5. Calculate Loss => Calculate Loss by using LiDAR for Ground Truth or by warping the left and right images. Recently, it is often seen that learning is done only with images.
Computing the Stereo Matching Cost with a Convolutional Neural Network(2015)
Improve accuracy by substituting multi-channel features by convolving the features of the right and left images instead of intensity
Spatial Pyramid Pooling in Deep Convolutional
Networks for Visual Recognition(2015)
You have to refer to a wider range of pixels to match nearby objects. The solution is to combine the fine and coarse resolutions of the feature map.
End-to-End Learning of Geometry and Context for Deep Stereo Regression(2017)
As before, the right and left images are convoluted using the same weight to generate a feature map (WxHxC). Create a feature map (DxWxHxC) by shifting the pixels of the left image horizontally from 0 to maxDisparity (arbitrary) with respect to the right image. Simply shift the pixel horizontally (width direction). By performing 3D Convolution and 3D Deconvolution at 1/2, 1/4, 1/8, 1/16, 1/32, you can learn rough features and detailed features. The output here is (DxHxW). The final Disparity is output by multiplying the one-dimensional matching value by Disparity and taking the weighted average. With softArgMin, it is possible to calculate Disparity with sub-pixel accuracy.
Self-Supervised Learning for Stereo Matching with Self-Improving Ability(2017)
Until now, I have obtained Disparity or Depth from LiDAR to find Loss. The density of LiDAR is coarser than that of the image, and since training is performed even in a system that does not use LiDAR, the image on the left is simulated by shifting the pixel by the estimated Disparity on the right. Loss can be defined by looking at the generated left image and the original left image by looking at SAD (intensity or RGB difference) or SSIM (structural similarity). If the Disparity can be estimated correctly, the warped image will be almost the same as the opposite image.
Recommended Posts