Local Similarity Pattern and Cost Self-Reassembling for Deep Stereo Matching Networks

Contributions

  • A novel pairwise feature LSP to extract structural information, which is beneficial for accurate matching especially when the illumination of the image pair is imbalanced
  • A novel disparity refinement method CSR (or DSR to save memory) to deal with outliers that are difficult to match, e.g. disparity discontinuities and occluded regions.

Dependencies:

Training on SceneFlow

python train.py --data_path (your Scene Flow data folder)

Finetuning on KITTI

python KITTI_ft.py --data_path (your KITTI training data folder) --load_path (the path of the model trained on SceneFlow)

GitHub

View Github