This is the code for the CVPR'19 paper "Learning Regularity in Skeleton Trajectories for Anomaly Detection in Videos".

First please create an appropriate environment using conda:

conda env create -f environment.yml

conda activate tbad

Due to space constraints in the Github repository, please download the data from the following link and place the data folder in this directory.

Link: trajectories

To evaluate pre-trained models run the evaluate.py script. Some examples:

python evaluate.py --gpu_ids 0 --gpu_memory 0.2 combined_models ./pretrained/CVPR19/ShanghaiTech/combined_model/_mp_Grobust_Lrobust_Orobust_concatdown_ ./data/HR-ShanghaiTech/testing/trajectories ./data/HR-ShanghaiTech/testing/frame_level_masks --video_resolution 856x480 --overlapping_trajectories

python evaluate.py --gpu_ids 0 --gpu_memory 0.2 combined_model ./pretrained/CVPR19/Avenue/combined_model/_mp_Grobust_Lrobust_Orobust_concatdown_/01_2018_11_13_06_36_20 ./data/HR-Avenue/testing/trajectories/01 ./data/HR-Avenue/testing/frame_level_masks/01 --video_resolution 640x360 --overlapping_trajectories

To train a model from scratch you should look up the model's configuration options using the option --help on the training.py script. Here is one example:

python train.py --gpu_ids 0 --gpu_memory 0.1 combined_model ./data/HR-ShanghaiTech/training/trajectories/00 --video_resolution 856x480 --message_passing --reconstruct_original_data --multiple_outputs --multiple_outputs_before_concatenation --input_length 12 --rec_length 12 --pred_length 6 --reconstruct_reverse --cell gru --global_hidden 8 --local_hidden 16 --output_activation linear --optimiser adam --learning_rate 0.001 --loss mse --epochs 5 --batch_size 256 --global_normalisation robust --local_normalisation robust --out_normalisation robust