Stereo Matching Algorithm Based on CGAN
Dons Wei, Han Liu · 2021 China Automation Congress (CAC) · 2021
Most of the current end-to-end stereo matching algorithms use 3D convolution, which leads to high computational cost. Therefore, this paper proposes a algorithm based on CGAN to solve this problem. Because the result of prediction is presented in the way of disparity map, the task of disparity prediction can be regarded as the task of image generation. The generator in CGAN converts the pairs of images into the disparity maps. The generated image or the ground truth will be input to the discriminator in CGAN to identify whether the input image is a generated sample or the ground truth. The errors calculated from the generated results and the discriminant results guide the training of the generator and the discriminator to form a adversarial process until both are optimal. Our proposed method does not use 3D convolution, which not only simplifies the realization process of stereo matching, but also reduces the calculation cost on the premise of ensuring certain accuracy. We evaluated our model on the Scene Flow dataset and achieved a good result by comparing different architectures.