Contrast and Mix (CoMix)

The repository contains the codes for the paper Contrast and Mix: Temporal Contrastive Video Domain Adaptation with Background Mixing part of Advances in Neural Information Processing Systems (NeurIPS) 2021.

Aadarsh Sahoo¹, Rutav Shah¹, Rameswar Panda², Kate Saenko^2,3, Abir Das¹

¹ IIT Kharagpur, ² MIT-IBM Watson AI Lab, ³ Boston University

[Paper] [Project Page]

Fig. Temporal Contrastive Learning with Background Mixing and Target Pseudo-labels. Temporal contrastive loss (left) contrasts a single temporally augmented positive (same video, different speed) per anchor against rest of the videos in a mini-batch as negatives. Incorporating background mixing (middle) provides additional positives per anchor possessing same action semantics with a different background alleviating background shift across domains. Incorporating target pseudo-labels (right) additionally enhances the discriminabilty by contrasting the target videos with the same pseudo-label as positives against rest of the videos as negatives.

Preparing the Environment

Conda

Please use the comix_environment.yml file to create the conda environment comix as:

conda env create -f comix_environment.yml

Pip

Please use the requirements.txt file to install all the required dependencies as:

pip install -r requirements.txt

Data Directory Structure

All the datasets should be stored in the folder ./data following the convention ./data/<dataset_name> and it must be passed as an argument to base_dir=./data/<dataset_name>.

UCF - HMDB

For ucf_hmdb dataset with base_dir=./data/ucf_hmdb the structure would be as follows:

.
├── ...
├── data
│   ├── ucf_hmdb
│   │   ├── ucf_videos
|   |   |   ├── <video-1>
|   |   |   |   ├── <frame-1>
|   |   |   |   ├── <frame-2>
|   |   |   |   ├── ...
|   |   |   ├── <video-2>
|   |   |   ├── ...
│   │   ├── hmdb_videos
|   |   ├── ucf_BG
|   |   └── hmdb_BG
│   └──
└──

Jester

For Jester dataset with base_dir=./data/jester the structure would be as follows

.
├── ...
├── data
│   ├── jester
|   |   ├── jester_videos
|   |   |   ├── <video-1>
|   |   |   |   ├── <frame-1>
|   |   |   |   ├── <frame-2>
|   |   |   |   ├── ...
|   |   |   ├── <video-2>
|   |   |   ├── ...
|   |   ├── jester_BG
|   |   |   ├── <video-1>
|   |   |   |   ├── <Backgorund frame>
|   |   |   ├── ...
└── └── └──

Epic-Kitchens

For Epic Kitchens dataset with base_dir=./data/epic_kitchens the structure would be as follows (we follow the same structure as in the original dataset) :

.
├── ...
├── data
│   ├── epic_kitchens
|   |   ├── epic_kitchens_videos
|   |   |   ├── train
|   |   |   |   ├── D1
|   |   |   |   |   ├── <video-1>
|   |   |   |   |   |   ├── <frame-1>
|   |   |   |   |   |   ├── <frame-2>
|   |   |   |   |   |   ├── ...
|   |   |   |   |   ├── <video-2>
|   |   |   |   |   ├── ...
|   |   |   |   ├── D2
|   |   |   |   └── D3
|   |   |   └── test
└── └── └── epic_kitchens_BG

For using datasets stored in some other directories, please pass the parameter base_dir accordingly.

Background Extraction using Temporal Median Filtering

Please refer to the folder ./background_extraction for the codes to extract backgrounds using temporal median filtering.

Data

All the required split files are provided inside the directory ./video_splits.

The official download links for the datasets used for this paper are: [UCF-101] [HMDB-51] [Jester] [Epic Kitchens]

Training CoMix

Here are some of the sample and recomended commands to train CoMix for the transfer task of:

UCF -> HMDB from UCF-HMDB dataset:

CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --manual_seed 1 --dataset_name UCF-HMDB --src_dataset UCF --tgt_dataset HMDB --batch_size 8 --model_root ./checkpoints_ucf_hmdb --save_in_steps 500 --log_in_steps 50 --eval_in_steps 50 --pseudo_threshold 0.7 --warmstart_models True --num_iter_warmstart 4000 --num_iter_adapt 10000 --learning_rate 0.01 --learning_rate_ws 0.01 --lambda_bgm 0.1 --lambda_tpl 0.01 --base_dir ./data/ucf_hmdb

S -> T from Jester dataset:

CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --manual_seed 1 --dataset_name Jester --src_dataset S --tgt_dataset T --batch_size 8 --model_root ./checkpoints_jester --save_in_steps 500 --log_in_steps 50 --eval_in_steps 50 --pseudo_threshold 0.7 --warmstart_models True --num_iter_warmstart 4000 --num_iter_adapt 10000 --learning_rate 0.01 --learning_rate_ws 0.01 --lambda_bgm 0.1 --lambda_tpl 0.1 --base_dir ./data/jester

D1 -> D2 from Epic-Kitchens dataset:

CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --manual_seed 1 --dataset_name Epic-Kitchens --src_dataset D1 --tgt_dataset D2 --batch_size 8 --model_root ./checkpoints_epic_d1_d2 --save_in_steps 500 --log_in_steps 50 --eval_in_steps 50 --pseudo_threshold 0.7 --warmstart_models True --num_iter_warmstart 4000 --num_iter_adapt 10000 --learning_rate 0.01 --learning_rate_ws 0.01 --lambda_bgm 0.01 --lambda_tpl 0.01 --base_dir ./data/epic_kitchens

For detailed description regarding the arguments, use:

python main.py --help

Citing CoMix

If you use codes in this repository, consider citing CoMix. Thanks!

@article{sahoo2021contrast,
  title={Contrast and Mix: Temporal Contrastive Video Domain Adaptation with Background Mixing},
  author={Sahoo, Aadarsh and Shah, Rutav and Panda, Rameswar and Saenko, Kate and Das, Abir},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}

srikanth-sfu/CoMix