HSTRNet-interp: Dual Camera Based High Spatio-Temporal Resolution Video Generation For Wide Area Surveillance

Introduction

This repository contains implementation of paper titled Dual Camera Based High Spatio-Temporal Resolution Video Generation For Wide Area Surveillance(IEEE). HSTRNet-interp is a deep-learning network capable of generating high spatiotemporal resolution videos from high spatial resolution - low frame rate and low spatial resolution - high frame rate feeds.

Getting Started

Run following commands to setup the environment and download weights to proper locations:

git clone https://github.com/umutsuluhan/HSTRNet-Interp.git
source create_env.sh

Testing

Run following command to test model on different datasets:

python3 scripts/eval.py --dataset_name <dataset_name> --dataset_path <dataset_path> --model_path <model_path>

Parameters are:

dataset_name	Vimeo	Vizdrone
dataset_path	Vimeo dataset path	pretrained/vimeo
dataset_name	Visdrone dataset path	pretrained/visdrone

Training

Run following command to train the model:

python3 scripts/train.py --data_root <dataset_path> --train_batch_size <training_batch_size> --val_batch_size <validation_batch_size> --epoch <epoch_count> --checkpoint <0,1> --checkpoint_start <checkpoint_epoch_number> --checkpoint_path <checkpoint_path>

• checkpoint flag is set to 0 if training is being started from scratch and 1 if training will continue from a checkpoint
• checkpoint_start is the epoch number to continue if checkpoint is set to 1

Citation

Please cite our paper in the following format if you use this codebase for academic purposes:

@inproceedings{suluhan2022dual,
  title={Dual Camera Based High Spatio-Temporal Resolution Video Generation For Wide Area Surveillance},
  author={Suluhan, H Umut and Ates, Hasan F and Gunturk, Bahadir K},
  booktitle={2022 18th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS)},
  pages={1--8},
  year={2022},
  organization={IEEE}
}