Codes for the paper Contrast and Mix: Temporal Contrastive Video Domain Adaptation with Background Mixing

Last update: Dec 10, 2022

Overview

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/ and it must be passed as an argument to base_dir=./data/.

UCF - HMDB

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

.
├── ...
├── data
│   ├── ucf_hmdb
│   │   ├── ucf_videos
|   |   |   ├── 
   
    
|   |   |   |   ├── 
    
     
|   |   |   |   ├── 
     
      
|   |   |   |   ├── ...
|   |   |   ├── 
      
       
|   |   |   ├── ...
│   │   ├── 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
|   |   |   ├── 
   
    
|   |   |   |   ├── 
    
     
|   |   |   |   ├── 
     
      
|   |   |   |   ├── ...
|   |   |   ├── 
      
       
|   |   |   ├── ...
|   |   ├── jester_BG
|   |   |   ├── 
       
         | | | | ├── 
        
          | | | ├── ... └── └── └──

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
|   |   |   |   |   ├── 
   
    
|   |   |   |   |   |   ├── 
    
     
|   |   |   |   |   |   ├── 
     
      
|   |   |   |   |   |   ├── ...
|   |   |   |   |   ├── 
      
       
|   |   |   |   |   ├── ...
|   |   |   |   ├── 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}
}

Codes for the paper Contrast and Mix: Temporal Contrastive Video Domain Adaptation with Background Mixing

Related tags

Overview

Contrast and Mix (CoMix)

Preparing the Environment

Conda

Pip

Data Directory Structure

UCF - HMDB

Jester

Epic-Kitchens

Background Extraction using Temporal Median Filtering

Data

Training CoMix

Citing CoMix

Owner

Computer Vision and Intelligence Research (CVIR)

Nonnegative spatial factorization for multivariate count data

Unofficial Pytorch Lightning implementation of Contrastive Syn-to-Real Generalization (ICLR, 2021)

An implementation demo of the ICLR 2021 paper Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks in PyTorch.

PyTorch Implementation of CycleGAN and SSGAN for Domain Transfer (Minimal)

State-to-Distribution (STD) Model

Composing methods for ML training efficiency

A collection of Jupyter notebooks to play with NVIDIA's StyleGAN3 and OpenAI's CLIP for a text-based guided image generation.

AdamW optimizer and cosine learning rate annealing with restarts

VOGUE: Try-On by StyleGAN Interpolation Optimization

Parallel Latent Tree-Induction for Faster Sequence Encoding

CS5242_2021 - Neural Networks and Deep Learning, NUS CS5242, 2021

DIVeR: Deterministic Integration for Volume Rendering

A programming language written with python

Underwater industrial application yolov5m6

This repository holds code and data for our PETS'22 article 'From "Onion Not Found" to Guard Discovery'.

4D Human Body Capture from Egocentric Video via 3D Scene Grounding

Multi-Stage Episodic Control for Strategic Exploration in Text Games

Python implementation of Bayesian optimization over permutation spaces.

Uni-Fold: Training your own deep protein-folding models.

Pytorch implementation of various High Dynamic Range (HDR) Imaging algorithms