On the Benefits of 3D Pose and Tracking for Human Action Recognition (CVPR 2023)

Code repository for the paper "On the Benefits of 3D Pose and Tracking for Human Action Recognition".
Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, Christoph Feichtenhofer, Jitendra Malik.

This code repository provides a code implementation for our paper LART (Lagrangian Action Recognition with Tracking), with installation, training, and evaluating on datasets, and a demo code to run on any youtube videos.

Installation

After installing the PyTorch 2.0 dependency, you can install our lart package directly as:

pip install git+https://github.com/brjathu/LART

Step-by-step instructions

conda create -n lart python=3.10
conda activate lart
pip install torch==2.0.0+cu117 torchvision==0.15.1+cu117 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu117
pip install -e .[demo]

If you only wants to train the model and not interested in running demo on any videos, you dont need to install packages rquired for demo code (pip install -e .).

Demo on videos

We have provided a colab notebook to run our model on any youtube video. Please see the demo notebook for details. Since colab has 15 Gb memory limit, the dev branch is running on half precision. If you want to run the demo on full precision, please clone the repo and run the demo locally.

Our Action recogition model uses PHALP to track people in videos, and then uses the tracklets to classify actions. pip install -e .[demo] will install nessory dependencies for running the demo code. Now, run demo.py to reconstruct, track and recognize humans in any video. Input video source may be a video file, a folder of frames, or a youtube link:

# Run on video file
python scripts/demo.py video.source="assets/jump.mp4"

# Run on extracted frames
python scripts/demo.py video.source="/path/to/frames_folder/"

# Run on a youtube link (depends on pytube working properly)
python scripts/demo.py video.source=\'"https://www.youtube.com/watch?v=xEH_5T9jMVU"\'

The output directory (./outputs by default) will contain a video rendering of the tracklets and a .pkl file containing the tracklets with 3D pose and shape and action labels. Please see the PHALP repository for details. The model for demo, uses MViT as a backend and a single person model. The demo code requires about 32 GB memory to run the slowfast code. [TODO: Demo with Hiera backend.]

Training and Evaluation

Download the datasets

We are releasing about 1.5 million tracks of people from Kinetics 400 and AVA datasets. Run the following command to download the data from dropbox and extract them to the data folder. This will download preprocessed data form ava_val, ava_train and kinetics_train datasets (preporcessed data for AVA-kinetics and multispots will be released soon). These tracjectories contains backbone features as well as ground-truth annotations and pseudo ground-truth annotations. For more details see DATASET.md

./scripts/download_data.sh

Train LART model

For single node training, run the following command. This will evaulate the model at every epochs and compute the mean average precision on the validation set.

# # LART full model. 
python lart/train.py -m \
--config-name lart.yaml \
trainer=ddp_unused \
task_name=LART \
trainer.devices=8 \
trainer.num_nodes=1 \

Evaluate LART model

First download the pretrained model by running the following command. This will download the pretrained model to ./logs/LART_Hiera/ folder.

./scripts/download_checkpoints.sh

Then run the following command to evaluate the model on the validation set. This will compute the mean average precision on the validation set (AVA-K evaluation will be released soon). For AVA, provided checkpoint should give 45.1 mAP.

# # LART full model. 
python lart/train.py -m \
--config-name lart.yaml \
trainer=ddp_unused \
task_name=LART_eval \
train=False \
trainer.devices=8 \
trainer.num_nodes=1 \
[email protected] \
configs.weights_path=logs/LART_Hiera/0/checkpoints/epoch_002-EMA.ckpt \

Please specify a different configs.weights_path to evaluate your own trained model. However every checkpoint will be evaluated during training, and results will be saved to ./logs/<MODEL_NAME>/0/results/ folder.

Results

The following table shows the performance of our model on AVA-2.2 dataset.

Model	Pretrain	mAP
SlowFast R101, 8×8	K400	23.8
SlowFast 16×8 +NL	K600	27.5
X3D-XL	K600	27.4
MViTv1-B-24, 32×3	K600	28.7
Object Transformer	K600	31.0
ACAR R101, 8×8 +NL	K700	33.3
MViT-L↑312, 40×3	IN-21K+K400	31.6
MaskFeat	K600	38.8
Video MAE	K400	39.5
Hiera	K700	43.3
LART - MViT	K400	42.3 (+2.8)
LART - Hiera	K400	45.1 (+2.5)

Citation

If you find this code useful for your research or the use data generated by our method, please consider citing the following paper:

@inproceedings{rajasegaran2023benefits,
  title={On the Benefits of 3D Pose and Tracking for Human Action Recognition},
  author={Rajasegaran, Jathushan and Pavlakos, Georgios and Kanazawa, Angjoo and Feichtenhofer, Christoph and Malik, Jitendra},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={640--649},
  year={2023}
}