PRBonn / segcontrast

MIT License
96 stars 13 forks source link

SegContrast

Paper | Video

Installing pre-requisites:

sudo apt install build-essential python3-dev libopenblas-dev

pip3 install -r requirements.txt

pip3 install torch ninja

Installing MinkowskiEngine with CUDA support:

pip3 install -U MinkowskiEngine==0.5.4 --install-option="--blas=openblas" -v --no-deps


Note: We have released a new representation learning method based on temporal associations (TARL) that achieves better performance than SegContrast. Besides, we have added a new branch tarl/fine-tuning to this repo with the hyperparameters used in the fine-tuning experiments reported in the TARL paper.

SegContrast with Docker

Inside the docker/ directory there is a Dockerfile to build an image to run SegContrast. You can build the image from scratch or download the image from docker hub by:

docker pull nuneslu/segcontrast:minkunet

Then start the container with:

docker run --gpus all -it --rm -v /PATH/TO/SEGCONTRAST:/home/segcontrast segcontrast /bin/zsh

Data Preparation

Download SemanticKITTI inside the directory ./Datasets/SemanticKITTI/datasets. The directory structure should be:

./
└── Datasets/
    └── SemanticKITTI
        └── dataset
          └── sequences
            ├── 00/           
            │   ├── velodyne/   
            |   |   ├── 000000.bin
            |   |   ├── 000001.bin
            |   |   └── ...
            │   └── labels/ 
            |       ├── 000000.label
            |       ├── 000001.label
            |       └── ...
            ├── 08/ # for validation
            ├── 11/ # 11-21 for testing
            └── 21/
                └── ...

Pretrained Weights

Reproducing the results

Run the following to start the pre-training:

python3 contrastive_train.py --use-cuda --use-intensity --segment-contrast --checkpoint segcontrast

The default parameters, e.g., learning rate, batch size and epochs are already the same as the paper.

After pre-training you can run the downstream fine-tuning with:

python3 downstream_train.py --use-cuda --use-intensity --checkpoint \
        segment_contrast --contrastive --load-checkpoint --batch-size 2 \
        --sparse-model MinkUNet --epochs 15

We provide in tools the contrastive_train.sh and downstream_train.sh scripts to reproduce the results pre-training and fine-tuning with the different label percentages shown on the paper:

For pre-training:

./tools/contrastive_train.sh

Then for fine-tuning:

./tools/downstream_train.sh

Finally, to compute the IoU metrics use:

./tools/eval_train.sh

Citation

If you use this repo, please cite as :

@article{nunes2022ral,
    author = {L. Nunes and R. Marcuzzi and X. Chen and J. Behley and C. Stachniss},
    title = {{SegContrast: 3D Point Cloud Feature Representation Learning through Self-supervised Segment Discrimination}},
    journal = {{IEEE Robotics and Automation Letters (RA-L)}},
    year = 2022,
    doi = {10.1109/LRA.2022.3142440},
    issn = {2377-3766},
    volume = {7},
    number = {2},
    pages = {2116-2123},
    url = {http://www.ipb.uni-bonn.de/pdfs/nunes2022ral-icra.pdf},
}