Home

Awesome

Pytorch Code for VideoLT

[Website][Paper]

Updates

<img src="https://github.com/17Skye17/VideoLT/blob/master/pics/concept.png" width="400" alt="concept">

Overview

VideoLT is a large-scale long-tailed video recognition dataset, as a step toward real-world video recognition. We provide VideoLT dataset and long-tailed baselines in this repo including:

Data Preparation

Please be aware that VideoLT is only for non-commercial use, please send us an e-mail: zxwu at fudan.edu.cn and agree to our license, then we will send back the download links to you. We provide raw videos(~1.7TB) and extracted features(~900GB in total, ~295GB for each).

To decompress the .tar.gz files, please use commands:

cat TSM-R50-feature.tar.gz.part* | tar zx 
cat ResNet50-feature.tar.gz.part* | tar zx
cat ResNet101-feature.tar.gz.part* | tar zx

For using extracted features, please modify dataset/dutils.py and set the correct path to features.

Model Zoo

The baseline scripts and checkpoints are provided in MODELZOO.md.

FrameStack

FrameStack is simple yet effective approach for long-tailed video recognition which re-samples training data at the frame level and adopts a dynamic sampling strategy based on knowledge learned by the network. The rationale behind FrameStack is to dynamically sample more frames from videos in tail classes and use fewer frames for those from head classes.

<img src="https://github.com/17Skye17/VideoLT/blob/master/pics/framestack.png" width="400" alt="framestack">

Usage

Requirement

pip install -r requirements.txt

Prepare Data Path

  1. Modify FEATURE_NAME, PATH_TO_FEATURE and FEATURE_DIM in dataset/dutils.py.

  2. Set ROOT in dataset/dutils.py to labels folder. The directory structure is:

    labels
    |-- count-labels-train.lst
    |-- test.lst
    |-- test_videofolder.txt
    |-- train.lst
    |-- train_videofolder.txt
    |-- val_videofolder.txt
    `-- validate.lst

Train

We provide scripts for training. Please refer to MODELZOO.md.

Example training scripts:

FEATURE_NAME='ResNet101'

export CUDA_VISIBLE_DEVICES='2'
python base_main.py  \
     --augment "mixup" \
     --feature_name $FEATURE_NAME \
     --lr 0.0001 \
     --gd 20 --lr_steps 30 60 --epochs 100 \
     --batch-size 128 -j 16 \
     --eval-freq 5 \
     --print-freq 20 \
     --root_log=$FEATURE_NAME-log \
     --root_model=$FEATURE_NAME'-checkpoints' \
     --store_name=$FEATURE_NAME'_bs128_lr0.0001_lateavg_mixup' \
     --num_class=1004 \
     --model_name=NonlinearClassifier \
     --train_num_frames=60 \
     --val_num_frames=150 \
     --loss_func=BCELoss \

Note: Set args.resample, args.augment and args.loss_func can apply multiple long-tailed stratigies.

Options:

    args.resample: ['None', 'CBS','SRS']
    args.augment : ['None', 'mixup', 'FrameStack']
    args.loss_func: ['BCELoss', 'LDAM', 'EQL', 'CBLoss', 'FocalLoss']

Test

We provide scripts for testing in scripts. Modify CKPT to saved checkpoints.

Example testing scripts:

FEATURE_NAME='ResNet101'
CKPT='VideoLT_checkpoints/ResNet-101/ResNet101_bs128_lr0.0001_lateavg_mixup/ckpt.best.pth.tar'

export CUDA_VISIBLE_DEVICES='1'
python base_test.py \
     --resume $CKPT \
     --feature_name $FEATURE_NAME \
     --batch-size 128 -j 16 \
     --print-freq 20 \
     --num_class=1004 \
     --model_name=NonlinearClassifier \
     --train_num_frames=60 \
     --val_num_frames=150 \
     --loss_func=BCELoss \

Citing

If you find VideoLT helpful for your research, please consider citing:

@InProceedings{Zhang_2021_ICCV,
    author    = {Zhang, Xing and Wu, Zuxuan and Weng, Zejia and Fu, Huazhu and Chen, Jingjing and Jiang, Yu-Gang and Davis, Larry S.},
    title     = {VideoLT: Large-Scale Long-Tailed Video Recognition},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {7960-7969}
}