Authors: Yijun Qian, Lijun Yu, Wenhe Liu, Alexander G Hauptmann

Email: [email protected]

This is the official implementation of Rethinking Zero-shot Action Recognition: Learning from Latent Atomic Actions (ECCV22). For more information please refer to our accepted paper in ECCV 2022.

The code is built with many libraries. If you encounter problems about the dependencies, please resort to their official sites for help. We have prepared the environment config file and suggest build the environment through ANACONDA.

cd JigsawNet
conda env create -f environment.yml 
conda activate zsl

To run our code, please download Kinetics Dataset and place Kinetics400 and Kinetics600 under the same $Data Path. Make sure the data structure is like:

• $Data Path
  • Kinetics400
    • train
    • validate
    • test
  • Kinetics600
    • videos
      • trn
      • val
      • tst

Meanwhile, please download the annotation folder and place them under ./datasets KineticsZSAR-TRAIN KineticsZSAR-TEST

To quickly evaluate our model, please download the weight. Then follow these steps:

mkdir logs
mkdir exps
python eval.py --dataset KEVAL \
     --arch R2plus1D-34 --text_model bert \
     --epochs 1 --batch-size 48 -j 48 --dropout 0.5 --gpus 0 1 2 3 \
     --npb --lr_scheduler --vmz_tune_last_k_layer 4 \
     --freeze_text_to 10 --bert_pooling avg --consist_loss --attn \
     --root_model ./exps --root_log ./logs --video_path $Data Path \
     --tune_from $PATH-TO-MODEL

To train the model, please follow these steps:

mkdir logs
mkdir exps
python train.py --dataset KineticsZSAR \
     --arch R2plus1D-34 --text_model bert \
     --gd 20 --lr 1e-5 --lr_steps 2 6 9 12  --epochs 20 \
     --batch-size 28 -j 28 --dropout 0.5 --gpus 0 1 2 3 \
     --root_model ./exps --root_log ./logs --video_path /mnt/data \
     --npb --lr_scheduler --vmz_tune_last_k_layer 4 \
     --freeze_text_to 10 --bert_pooling avg  --attn --consist_loss

Please read the LICENSE before using.