Awesome
SENSE: a Shared Encoder for Scene Flow Estimation
PyTorch implementation of our ICCV 2019 Oral paper SENSE: A Shared Encoder for Scene-flow Estimation.
<p align="center"> <img src="sense.png" width="500" /> </p>License
Copyright (C) 2019 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode).
Requirements
- Python (tested with Python3.6.10)
- PyTorch (tested with 1.3.0)
- SynchronizedBatchNorm (borrowed from https://github.com/CSAILVision/semantic-segmentation-pytorch)
- tensorboardX
- tqdm
- OpenCV
- scipy
- numpy
It is strongly recommended to use a conda environment to install all the dependencies. Simply run sh scripts/install.sh to install all dependencies and also compile the correlation package.
All experiments were conducted on 8 2080ti GPUs (each with 11GB memory) or 2 M40 GPUs (each with 24GB memory).
In our original implementation, we used a C++ implementation for the cost volume computation for both optical flow and stereo disparity estimations. But the C++ implementatyion strictly requires a PyTorch version of 0.4.0. In this relased version, we switched to use the implemtnation provided at https://github.com/NVIDIA/flownet2-pytorch. We use this implementation for stereo disparity estimation, although it only supports cost volume computation for optical flow (searching for correspondence in a local 2D range). Please consult our paper if you are interested in the running time and GPU memory usuage.
Quick Start
First run sh scripts/download_pretrained_models.sh to download pre-trained models. Run python tools/demo.py then for a quick demo.
Run sh scripits/make_kitti2015_submission.sh to generate results that can be submitted to the online KITTI scene flow estimation benchmark. You should be able to get the following results.
| Error | D1-bg | D1-fg | D1-all | D2-bg | D2-fg | D2-all | Fl-bg | Fl-fg | Fl-all | SF-bg | SF-fg | SF-all |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| All/Est | 2.07 | 3.01 | 2.22 | 4.90 | 10.83 | 5.89 | 7.30 | 9.33 | 7.64 | 8.36 | 15.49 | 9.55 |
Training
See TRAINING.md for details.
Citation
If you find SENSE useful for your research, please consider citing it.
@InProceedings{jiang2019sense,
author = {Jiang, Huaizu and Sun, Deqing and Jampani, Varun and Lv, Zhaoyang and Learned-Miller, Erik and Kautz, Jan},
title = {SENSE: A Shared Encoder Network for Scene-Flow Estimation},
booktitle = {ICCV},
year = {2019}
}