This repo is the offical implementation of this paper.
This dataset is proposed by Sultani et al. To run this repo, simply download the dataset through the link and unzip to desired location.
The download link of this dataset is proposed by Liu et al. In addition, because the dataset is proposed for unsupervised learning task, we used the new split by GCN-Anomaly. The split file should be put in the SHtech dataset folder.
- update the dataset folder path in src/config.py
- download the pretrained weights (Sports1M) of C3D from here and put into models/
- convert video to frames
python preprocess/videotoframe_UCF.py [--reset] #for UCFCrime
python preprocess/videotoframe_SH.py [--reset] #for ShanghaiTech
[--reset]: clean all genearted folders
- slice the frames into clips In our setting, all videos are sliced into 32 segments and each segment are sampled into 16 frames.
python preprocess/clip_UCF.py #for UCFCrime
python preprocess/clip_SH.py # for ShanghaiTech
python main.py [--dataset] [--savelog] [--name] [--note] [--model_path] [--attention_type] [--gpus] [--epoch] [--lr] [--batch_size]
[--dataset] (required str): UCF or SH
[--savelog] (store_true): The log file would be saved as {dataset_folder}/log/{name}.log and the trained model weights would be saved in {dataset_folder}/log/{name}/. The default name is the datetime.
[--name] (str): set the name of logfile
[--note] (str): available to add some notes in logfile.
[--model_path] (Path): Trained model weights' path if desired to resume training.
[--attention type] (str): gated or normal. The type of attention mechanism in TACnet.
[--gpus] (str): choose the training GPU. This doesn't support for multi GPUs training.
python predict.py [--dataset] [--model_path] [--load_pretrain] [--gpus] [--p_graph] [--c_grpah] [--attn_graph]
[--dataset] (required str): UCF or SH
[--model_path] (Path): Trained model weights' path
[--load_pretrain] (store_true): use the pretrained C3D weights or fine-tuned C3D weights.
[--p_graph] (store_true): output the performance graph of each testcase to {dataset_folder}/image/{name}/performance/
[--c_graph] (store_true): output the TSNE visualized cluster graph of each testcase to {dataset_folder}/image/{name}/cluster/
[--attn_graph] (store_true): output the visualization of attention weights of each testcase to {dataset_folder}/image/{name}/attn/
