FutureFramePrediction.pytorch

Introduction

This is a PyTorch implementation of Future Frame Prediction for Anomaly Detection – A New Baseline.

Input image	result

Preresquisties

• PyTorch
• tqdm
• Albumentations
• NumPy
• imageio

How to train

• For training, you should get a sequence of video frames which contains 'normal' situations.
• It is a good way to see UCSD Anomaly Detection Dataset to get more details.
• Put a sequence and set your training phase at train.py

class args():
    # Dataset setting
    channels = 3
    size = 256
    frames_dir = 'datasets/train'
    time_steps = 5

where time_steps denotes each of size using the sliding window technique.

• Download a pretrained FlowNet model from Google Drive and write its location as follows

class args():
    # optical flow setting
    flownet_pretrained = 'pretrained/FlowNet2-SD.pth'

• Run the command below

python train.py

How to get anomalies

• Just run the command below

python demo.py

• Then, you can get GIF files(results/heatmap.gif) for anomalies.

To-do list

• Add evaluation codes

Notes

• For architectures in this repository, I refered to great works such FlowNet, UNet and PatchDiscriminator
• You can see same results with TensorFlow in official repository, and another PyTorch implementation