Awesome

CYBORGS

This repository contains the source code for our paper:

CYBORGS: Contrastively Bootstrapping Object Representations by Grounding in Segmentation
ECCV 2022
Renhao Wang, Hang Zhao, Yang Gao

Requirements

Code

Tested on Ubuntu 18.04 with CUDA 11.3. We include an environment file for fast setup. We also borrow heavily from the VirTex code base, which can be installed as per their docs.

# Example environment setup
conda env create -f cyborgs.yaml
conda activate cyborgs
pip install torch==1.10.0+cu113 torchaudio==0.10.0+cu113 torchvision==0.11.1+cu113 -f https://download.pytorch.org/whl/torch_stable.html
git clone https://www.github.com/kdexd/virtex
pip install -r virtex/requirements.txt
pip install git+git://github.com/facebookresearch/fvcore.git#egg=fvcore
pip install git+git://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI
cd virtex
python setup.py develop

Data

All datasets follow the setup and structure delineated in the VirTex code base.

Demos

Download a pretrained CYBORGS checkpoint and place the weights under /path/to/virtex/ckpts/cyborgs. Our downstream finetuning follows Detectron2-like configs provided in the VirTex code base. For example, to reproduce our results from the paper on PASCAL VOC 2007+2012 object detection:

cd /path/to/virtex
python scripts/eval_detectron2.py \
    --config ckpts/cyborgs/pretrain_config.yaml \
    --d2-config configs/detectron2/voc_det_default_init_24k.yaml \
    --checkpoint-path ckpts/cyborgs/cyborgs_frozen_800.pth \
    --weight-init torchvision \
    --num-gpus-per-machine 8 \
    --cpu-workers 2 \
    --serialization-dir ckpts/cyborgs/voc_det_500000 \
    --checkpoint-every 25000

For more details, kindly refer to the docs.

Training

To perform CYBORGS pretraining with the default (best) settings:

python main_cyborgs.py \
    --dist_url 'tcp://localhost:10020' \
    --num_gpus_per_machine 8 \
    --data_dir /path/to/coco \
    --mask_dir /path/to/bootstrapped/masks \
    --epochs 800 \
    --warmup_epoch 10 \
    --lr_scheduler step \
    --lr_decay_rate 0.1 \
    --lr_decay_epochs 768 784 \
    --batch_size 16 \
    --num_classes 81 \
    --min_clusters 2 \
    --max_clusters 256 \
    --boot_freq 100 \
    --num_workers 16 \
    --layer_name layer2.2 \
    --imgs_per_batch 16 \
    --num_proc 48 \
    --output_dir save/detcon_boot_100

Evaluation

The following downstream evaluation tasks use the VirTex code base as well:

• PASCAL VOC 2007 Linear Classification
• PASCAL VOC 2007+2012 Object Detection
• ImageNet Linear Classification
• COCO Instance Segmentation

Acknowledgements

This work is all about bootstrapping self-supervised representations, and our contributions are very much bootstrapped (in a self-supervised way!) from a wonderful community of open source resources. We owe a great debt of gratitude to Karan Desai for his wonderfully written repository VirTex, and also to Olivier Henaff and his co-authors for DetCon, which heavily inspired the core self-supervised learning component of our algorithm.

Citation

If you find our work or code helpful, please consider citing us:

@inproceedings{wang2022cyborgs,
  title={CYBORGS: Contrastively Bootstrapping Object Representations by Grounding in Segmentation},
  author={Wang, Renhao and Zhao, Hang and Gao, Yang},
  booktitle={European conference on computer vision},
  year={2022}
}