Official Pytorch implementation for AAAI2021 paper "RSPNet: Relative Speed Perception for Unsupervised Video Representation Learning"
All dependencies can be installed using pip:
python -m pip install -r requirements.txtOur experiments run on Python 3.7 and PyTorch 1.6. Other versions should work but are not tested.
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_VIDEOSBe warned, this will use all your CPU and will take several hours (on our Intel E5-2630 *2 workstation) to complete.
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.
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_subsetThis 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.
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.jsonnetYou can use kinetics100 dataset for training by editing config/pretrain/moco-train-base.jsonnet (line 13)
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| Architecture | Pre-trained dataset | Pre-training epoch | Pre-trained model | Acc. on UCF101 | Acc. on HMDB51 |
|---|---|---|---|---|---|
| S3D-G | Kinetics-400 | 1000 | Download link | 93.7 | 64.7 |
| S3D-G | Kinetics-400 | 200 | Download link | 89.9 | 59.6 |
| R(2+1)D | Kinetics-400 | 200 | Download link | 81.1 | 44.6 |
| ResNet-18 | Kinetics-400 | 200 | Download link | 74.3 | 41.8 |
| C3D | Kinetics-400 | 200 | Download link | 76.7 | 44.6 |
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
| Architecture | k=1 | k=5 | k=10 | k=20 | k=50 |
|---|---|---|---|---|---|
| C3D | 36.0 | 56.7 | 66.5 | 76.3 | 87.7 |
| ResNet-18 | 41.1 | 59.4 | 68.4 | 77.8 | 88.7 |
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
-
DECORDError cannot find video stream with wanted index: -1
Some video from Kinetics dataset does not contain a valid video stream for some unknown reason. To filter them out, run
python utils/verify_video.py PATH/TO/VIDEOS, then copy the output to theblacklistconfig inconfig/dataset/kinetics{400,100}.libsonnet. You need to haveffmpeginstalled.
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}
}
For any question, please file an issue or contact
Peihao Chen: phchencs@gmail.com
Deng Huang: im.huangdeng@gmail.com