StereoNet: Guided Hierarchical Refinement for Real-Time Edge-Aware Depth prediction model in pytorch. ECCV2018
If you want to communicate with me about the StereoNet, please concact me without hesitating. My email:
Now, my model's speed can achieve 25 FPS on 540*960 img with the best result of 1.87 EPE_all with 16X multi model, 1.95 EPE_all with 16X single model on sceneflow dataset by end-to-end training. the following are the side outputs and the prediction example
- KITTI2015 submission: http://www.cvlibs.net/datasets/kitti/eval_scene_flow_detail.php?benchmark=stereo&result=94cf8e1c4f9d61df98c8e8cc8c726a34360208be ranking 179
- refercence[1]
If you find our work useful in your research, please consider citing:
@inproceedings{khamis2018stereonet, title={Stereonet: Guided hierarchical refinement for real-time edge-aware depth prediction}, author={Khamis, Sameh and Fanello, Sean and Rhemann, Christoph and Kowdle, Adarsh and Valentin, Julien and Izadi, Shahram}, booktitle={Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany}, pages={8--14}, year={2018} }
I implement the real-time stereo model according to the StereoNet model in pytorch. The speed can reach 30FPS with top performance. The speed can reach 60FPS with lower performance.
Method | EPE_all on sceneflow dataset | EPE_all on kitti2012 dataset | EPE_all on kitti2015 dataset |
---|---|---|---|
ours(8X single) | stage0:2.26 stage1:1.38 | ||
Reference[1] | stage1: 1.525 |
- Our code is released under MIT License (see LICENSE file for details).
- python3.6
- pytorch0.4
- run main8Xmulti.py
- finetune the performance beating the original paper.
- Do not design massive deep networks with multiple stages to improve kitti by 1%(no meaning doing this)
- Use metrics that matter for visual navigation (hint: not L1 depth error)
- ...
- Sceneflow pretrain weight https://drive.google.com/open?id=1bSwewxrRfmFCxZDyAtyYyQQiw05nSFI8.
- Thanks to Sameh Khamis' help