code-future-frame-prediction's Introduction

Future Frame Prediction

Pytorch implementation of video anomaly detection paper for CVPR 2018: Future Frame Prediction for Anomaly Detection – A New Baseline.
Most codes were obtained from the following GitHub page: [Link]
To understand the code, please refer to the Google Colab page (Korean): [Link]

I only trained the ped2 dataset, the result:

AUC	USCD Ped2
original implementation	95.4%
this implementation	95.5%

The network pipeline.

Environments

PyTorch >= 1.1.
Python >= 3.6.
opencv
sklearn
Other common packages.

Prepare

Download the ped2 dataset and put it under the data folder.

USCD Ped2
Google Drive

Download the FlowNet2-SD weight and put it under the flownet/pretrained folder.

FlowNet2-SD.pth
Google Drive

Download the trained model and put it under the weights folder.

best_model_ped2.pth
Google Drive

Train

# default option.
python train.py --dataset=ped2 
# change 'seed'.
python train.py --dataset=ped2 --manualseed=50
# change 'max iteration'.
python train.py --dataset=ped2 --iters=60000
# change 'model save interval'.
python train.py --dataset=ped2 --save_interval=10000
# change 'validation interval'.
python train.py --dataset=ped2 --val_interval=1000
# Continue training with latest model
python train.py --dataset=ped2 --resume=latest_ped2

Evalution

# default option.
python eval.py --dataset=ped2 --trained_model=best_model_ped2
# change 'show heatmap'.
python eval.py --dataset=ped2 --trained_model=best_model_ped2 --show_heatmap=True
# change 'show roc_curve'.
python eval.py --dataset=ped2 --trained_model=best_model_ped2 --show_curve=True