edwardyang12 / cascade-stereo

cascade-stereo

Casacde-Stereo on Messytable Dataset

Installation

Docker Image

isabella98/cascade-stereo-image:latest

Optional: Apex (not tested)

Install apex to enable synchronized batch normalization:

git clone https://github.com/NVIDIA/apex.git
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Train

./CasStereoNet/scripts/messytable_remote.sh $NUM_OF_GPUS $OUTPUT_PATH

Example: Use ./CasStereoNet/scripts/messytable_remote.sh 2 ./train_7_28/debug to train your model with 2 GPUs.

I have simplified the argparser part in main.py, it now only contains some basic args. You can use --gaussian-blur and --color-jitter in your shell file ( e.g. ./CasStereoNet/scripts/messytable_remote.sh) to enable gaussian blur and color jitter during training.

If you want to change some other args such as model parameters, (e.g. lr, batch_size), please go to CasStereoNet/configs/remote_train_config.yaml, it contains some parameters that we don't need to change often. If your dataset folder is not mounted as /cephfs, please go to the this file to change the path to your dataset. The parameters of data augmentation are also included in this file.

Please use different $OUTPUT_PATH for each job or you can use sub directories such as train_7_28/dataaug-0/ to get more organized.

Test

python ./CasStereoNet/test_on_sim_real.py --config-file ./CasStereoNet/configs/remote_test_config.yaml --model $PATH_TO_YOUR_MODEL --annotate $EXPERIMENT_ANNOTATION --exclude-bg

Use --exlude-bg if you want to exclude background when calculating the error metric.

Use --onreal if you are testing on real dataset, omit if you want to test on sim dataset.

Use --debug if you want to load less data (10) and see quick result.

Example:

python CasStereoNet/test_on_sim_real.py --config-file ./CasStereoNet/configs/remote_test_config.yaml --model /isabella-fast/Cascade-Stereo/outputs/7_19_dataaug-0/checkpoint_000006.ckpt --annotate test_7_28 --exclude-bg --onreal

Reference

@inproceedings{gu2019cas,
  title={Cascade Cost Volume for High-Resolution Multi-View Stereo and Stereo Matching},
  author={Gu, Xiaodong and Fan, Zhiwen and Zhu, Siyu and Dai, Zuozhuo and Tan, Feitong and Tan, Ping},
  journal={arxiv preprint arXiv:1912.06378},
  year={2019}
}