Learning to Adapt to Unseen Abnormal Activities under Weak Supervision

This repo is official PyTorch implementation of Learning to Adapt to Unseen Abnormal Activities under Weak Supervision (ACCV 2020).

Jaeyoo Park, Junha Kim, Bohyung Han

Data

• Download following data link and unzip under your $DATA_ROOT_DIR.
• You can set 'data_root_dir' as an argument in 'options.py'.
• We extract I3D features from raw UCF-Crime videos.
• We follow this to make video features into 32 segment features.
• GT_anomaly.pkl: Temporal annotations for all videos.
• exclustion.pkl: We find some of duplicate videos (e.g. same videos but different video name)
• frames.pkl: Number of frames for all videos

You need to follow directory structure of dataset as below.

{$DATA_ROOT_DIR}
|-- {$DATASET NAME}
|   |-- pkl_files
|   |-- {all_rgbs}  
|   |   |-- {$CLASS_NAME}  
|   |   |-- |-- video feature files (.npy)  
|   |-- {all_flows}  
|   |   |-- same structures as {all_rgbs}  
|   |-- {splits}  

For details, please check the downloaded data.

Run

• 'seed' is used for selecting target class (e.g. 1 for Abuse) of UCF-Crime dataset
• All arguments are in options.py.
• Simple running command is as follows.

1. pretrain: python main.py --mode pretrain --dataset $DATASET_NAME --seed $CLASS_NUM --save_chpt
2. meta-train: python main.py --mode meta_train --dataset $DATASET_NAME --seed $CLASS_NUM --save_chpt
3. meta-test

• Scratch: python main.py --mode eval --dataset $DATASET_NAME --seed $CLASS_NUM
• Pretrain: python main.py --mode eval --dataset $DATASET_NAME --seed $CLASS_NUM --chpt $NAME_OF_CHECKPOINT_BY_PRETRAIN
• Meta-train: python main.py --mode eval --dataset $DATASET_NAME --seed $CLASS_NUM --chpt $NAME_OF_CHECKPOINT_BY_METATRAIN --sampling
  • For meta-test, chpt format is like '{}epochs_exp0_seed1_lr1e-5_split1.pkl'.

Reference

@InProceedings{park2020learning,  
author = {Park, Jaeyoo, Kim, Junha, and Han, Bohyoung},  
title = {Learning to Adapt to Unseen Abnormal Activities under Weak Supervision},  
booktitle = {Asian Conference on Computer Vision (ACCV)},  
year = {2020}  
}