Event-driven Video Deblurring via Spatio-Temporal Relation-Aware Network

Authors: Chengzhi Cao, Xueyang Fu, Yurui Zhu, Gege Shi, Zheng-Jun Zha

IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that our STRA significantly outperforms several competing methods, e.g., on the HQF dataset, our network achieves up to 1.3 d B in terms of PSNR over the most advanced method.
Researcher Affiliation Academia University of Science and Technology of China, China chengzhicao@mail.ustc.edu.cn, xyfu@ustc.edu.cn, {zyr, sgg19990910}@mail.ustc.edu.cn, zhazj@ustc.edu.cn
Pseudocode No The paper provides network diagrams and mathematical equations but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes The code is available at https://github.com/Chengzhi-Cao/STRA.
Open Datasets Yes Our Spatio-Temporal Relation-Aware network (STRA) is trained based on the benchmark Go Pro dataset [Nah et al., 2017], composed of synthetic events, 2,103 pairs of blurring frames and sharp clear ground-truth frames. ... For evaluation in real-world events, we utilize HQF dataset [Stoffregen et al., 2020], including both real-world events and ground-truth frames captured from a DAVIS240C [Brandli et al., 2014]
Dataset Splits No The paper mentions training on the Go Pro dataset (2,103 pairs) and testing on Go Pro testing datasets (1,111 pairs) and HQF dataset, but does not explicitly describe a validation dataset or its split.
Hardware Specification Yes Our network is implemented using Pytorch on a single NVIDIA RTX 2080Ti GPU.
Software Dependencies No The paper states 'implemented using Pytorch' but does not specify a version number for PyTorch or any other software libraries used.
Experiment Setup Yes In the training process, we randomly cropped the sampled frames with the size of 256 256. For data augmentation, each patch was horizontally flipped with the probability of 0.5. We use a batch size of 8 training pairs and ADAM optimizer [Kingma and Ba, 2017] with parameter β1 = 0.9, β2 = 0.999. The maximum training epoch is set to 200, with the initial learning rate 10 4, then decays by 25% every 50 epochs.