Home

Awesome

RSPNet

Official Pytorch implementation for AAAI2021 paper "RSPNet: Relative Speed Perception for Unsupervised Video Representation Learning"

[Supplementary Materials]

Getting Started

Install Dependencies

All dependencies can be installed using pip:

python -m pip install -r requirements.txt

Our experiments run on Python 3.7 and PyTorch 1.6. Other versions should work but are not tested.

Transcode Videos (Optional)

This step is optional but will increase the data loading speed dramatically.

We decode the videos on the fly while training so we don't need to split frames. This makes disk IO a lot faster but increases CPU usage. This transcode step aims at reducing CPU consumed by decoding by 1) lower video resolution. 2) add more key frames.

To perform transcode, you need to have ffmpeg installed, then run:

python utils/transcode_dataset.py PATH/TO/ORIGIN_VIDEOS PATH/TO/TRANSCODED_VIDEOS

Be warned, this will use all your CPU and will take several hours (on our Intel E5-2630 *2 workstation) to complete.

Prepare Datasets

Your are expected to prepare date for pre-training (Kinetics-400 dataset) and fine-tuning (UCF101, HMDB51 and Something-something-v2 datasets). To let the scripts find datasets on your system, the recommended way is to create symbolic links in ./data directory to the actual path. We found this solution flexible.

The expected directory hierarchy is as follow:

├── data
│   ├── hmdb51
│   │   ├── metafile
│   │   │   ├── brush_hair_test_split1.txt
│   │   │   └── ...
│   │   └── videos
│   │       ├── brush_hair
│   │       │   └── *.avi
│   │       └── ...
│   ├── UCF101
│   │   ├── ucfTrainTestlist
│   │   │   ├── classInd.txt
│   │   │   ├── testlist01.txt
│   │   │   ├── trainlist01.txt
│   │   │   └── ...
│   │   └── UCF-101
│   │       ├── ApplyEyeMakeup
│   │       │   └── *.avi
│   │       └── ...
│   ├── kinetics400
│   │   ├── train_video
│   │   │   ├── answering_questions
│   │   │   │   └── *.mp4
│   │   │   └── ...
│   │   └── val_video
│   │       └── (same as train_video)
│   ├── kinetics100
│   │   └── (same as kinetics400)
│   └── smth-smth-v2
│       ├── 20bn-something-something-v2
│       │   └── *.mp4
│       └── annotations
│           ├── something-something-v2-labels.json
│           ├── something-something-v2-test.json
│           ├── something-something-v2-train.json
│           └── something-something-v2-validation.json
└── ...

Alternatively, you can change the path in config/dataset to match your system.

Build Kinetics-100 dataset (Optional)

Some of our ablation study experiments use the Kinetics-100 dataset for pre-training. This dataset is built by extract 100 classes from Kinetics-400, which has the smallest file size on the train set.

If you have Kinetics-400 available, you can build Kinetics-100 by:

python -m utils.build_kinetics_subset

This script will create symbolic links instead of copy data. It is expected to complete in a minute.

We have included a pre-built one at data/kinetics100_links and created the symbolic link data/kinetics100 that related to it. You need to have data/kinetics400 available at runtime.

Pre-training on Pretext Tasks

Now you have set up the environment. Run the following command to pre-train your models on pretext tasks.

export CUDA_VISIBLE_DEVICES=0,1,2,3
# Architecture: C3D
python pretrain.py -e exps/pretext-c3d -c config/pretrain/c3d.jsonnet
# Architecture: ResNet-18
python pretrain.py -e exps/pretext-resnet18 -c config/pretrain/resnet18.jsonnet
# Architecture: S3D-G
python pretrain.py -e exps/pretext-s3dg -c config/pretrain/s3dg.jsonnet
# Architecture: R(2+1)D
python pretrain.py -e exps/pretext-r2plus1d -c config/pretrain/r2plus1d.jsonnet

You can use kinetics100 dataset for training by editing config/pretrain/moco-train-base.jsonnet (line 13)

<!-- ```json dataset: kinetics400, // or kinetics100 ``` -->

Action Recognition

After pre-trained on pretext tasks, these models are fine-tuned to perform action recognition task on UCF101, HMDB51 and Something-something-v2 datasets.

export CUDA_VISIBLE_DEVICES=0,1
# Dataset: UCF101
#     Architecture: C3D ACC@1=76.71%
python finetune.py -c config/finetune/ucf101_c3d.jsonnet \
                   --mc exps/pretext-c3d/model_best.pth.tar \
                   -e exps/ucf101-c3d
#     Architecture: ResNet-18 ACC@1=74.33%
python finetune.py -c config/finetune/ucf101_resnet18.jsonnet \
                   --mc exps/pretext-resnet18/model_best.pth.tar \
                   -e exps/ucf101-resnet18
#     Architecture: S3D-G ACC@1=89.9%
python finetune.py -c config/finetune/ucf101_s3dg.jsonnet \
                   --mc exps/pretext-s3dg/model_best.pth.tar \
                   -e exps/ucf101-s3dg
#     Architecture: R(2+1)D ACC@1=81.1%
python finetune.py -c config/finetune/ucf101_r2plus1d.jsonnet \
                   --mc exps/pretext-r2plus1d/model_best.pth.tar \
                   -e exps/ucf101-r2plus1d

# Dataset: HMDB51
#     Architecture: C3D ACC@1=44.58%
python finetune.py -c config/finetune/hmdb51_c3d.jsonnet \
                   --mc exps/pretext-c3d/model_best.pth.tar \
                   -e exps/hmdb51-c3d
#     Architecture: ResNet-18 ACC@1=41.83%
python finetune.py -c config/finetune/hmdb51_resnet18.jsonnet \
                   --mc exps/pretext-resnet18/model_best.pth.tar \
                   -e exps/hmdb51-resnet18
#     Architecture: S3D-G ACC@1=59.6%
python finetune.py -c config/finetune/hmdb51_s3dg.jsonnet \
                   --mc exps/pretext-s3dg/model_best.pth.tar \
                   -e exps/hmdb51-s3dg
#     Architecture: R(2+1)D ACC@1=44.6%
python finetune.py -c config/finetune/hmdb51_r2plus1d.jsonnet \
                   --mc exps/pretext-r2plus1d/model_best.pth.tar \
                   -e exps/hmdb51-r2plus1d

# Dataset: Something-something-v2
#     Architecture: C3D ACC@1=47.76%
python finetune.py -c config/finetune/smth_smth_c3d.jsonnet \
                   --mc exps/pretext-c3d/model_best.pth.tar \
                   -e exps/smthv2-c3d
#     Architecture: ResNet-18 ACC@1=44.02%
python finetune.py -c config/finetune/smth_smth_resnet18.jsonnet \
                   --mc exps/pretext-resnet18/model_best.pth.tar \
                   -e exps/smthv2-resnet18
#     Architecture: S3D-G ACC@1=55.03%
python finetune.py -c config/finetune/smth_smth_s3dg.jsonnet \
                   --mc exps/pretext-s3dg/model_best.pth.tar \
                   -e exps/smthv2-s3dg

Results and Pre-trained Models

ArchitecturePre-trained datasetPre-training epochPre-trained modelAcc. on UCF101Acc. on HMDB51
S3D-GKinetics-4001000Download link93.764.7
S3D-GKinetics-400200Download link89.959.6
R(2+1)DKinetics-400200Download link81.144.6
ResNet-18Kinetics-400200Download link74.341.8
C3DKinetics-400200Download link76.744.6

Video Retrieval

The pretrained model can also be used in searching relevant videos based on the given query video.

export CUDA_VISIBLE_DEVICES=0 # use single GPU 
python retrieval.py -c config/retrieval/ucf101_resnet18.jsonnet \
                    --mc exps/pretext-resnet18/model_best.pth.tar \
                    -e exps/retrieval-resnet18    

The video retrieval result in our paper

Architecturek=1k=5k=10k=20k=50
C3D36.056.766.576.387.7
ResNet-1841.159.468.477.888.7

Visualization

We further visualize the region of interest (RoI) that contributes most to the similarity score using the class activation map (CAM) technique.

export CUDA_VISIBLE_DEVICES=0,1
python visualization.py -c config/pretrain/s3dg.jsonnet \
                        --load-model exps/pretext-s3dg/model_best.pth.tar \
                        -e exps/visual-s3dg \
                        -x '{batch_size: 1}'

The cam visualization results will be plotted in png files like

<div align="center"> <img src=resources/visualization.png width=50% /> </div>

Troubleshoot

Citation

Please cite the following paper if you feel RSPNet useful to your research

@InProceedings{chen2020RSPNet,
author = {Peihao Chen, Deng Huang, Dongliang He, Xiang Long, Runhao Zeng, Shilei Wen, Mingkui Tan, and Chuang Gan},
title = {RSPNet: Relative Speed Perception for Unsupervised Video Representation Learning},
booktitle = {The AAAI Conference on Artificial Intelligence (AAAI)},
year = {2021}
}

Contact

For any question, please file an issue or contact

Peihao Chen: phchencs@gmail.com
Deng Huang: im.huangdeng@gmail.com