Event-Enhanced Blurry Video Super-Resolution
Authors: Dachun Kai, Yueyi Zhang, Jin Wang, Zeyu Xiao, Zhiwei Xiong, Xiaoyan Sun
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive evaluations demonstrate that Ev-Deblur VSR establishes a new state-of-the-art performance on both synthetic and real-world datasets. Notably, on real data, our method is +2.59d B more accurate and 7.28 faster than the recent best BVSR baseline FMA-Net. |
| Researcher Affiliation | Academia | 1Mo E Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China 2Institute of Artificial Intelligence, Hefei Comprehensive National Science Center 3National University of Singapore |
| Pseudocode | No | The paper describes the architecture and processes of the Ev-Deblur VSR network, RFD module, and HDA module using descriptive text and figures, but it does not provide any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://github.com/Dachun Kai/Ev-Deblur VSR |
| Open Datasets | Yes | We use two widely-used datasets for training: Go Pro (Nah, Hyun Kim, and Mu Lee 2017) and BSD (Zhong et al. 2020). [...] For this, we use the recently published event-based motion deblurring dataset NCER (Cho et al. 2023) |
| Dataset Splits | Yes | The Go Pro dataset [...] contains 22 videos for training and 11 for testing. The BSD dataset [...] includes 60 sequences for training and 20 for testing. [...] NCER (Cho et al. 2023), which includes 27 videos for training (a total of 2,583 frames) and 16 videos for testing (1,454 frames). |
| Hardware Specification | Yes | The entire training process runs on 8 NVIDIA RTX4090 GPUs and takes about four days per dataset to converge. |
| Software Dependencies | No | The paper does not explicitly mention specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | when training, we use 15 frames as inputs, set the mini-batch size to 8, and center-crop the input frames size and event voxels size as 64 64. We use random horizontal and vertical flips to augment the data. On the three datasets mentioned above, we first train the model on Go Pro for 300K iterations using the Adam optimizer and Cosine Annealing scheduler. For the experiments on BSD, we fine-tune the model trained on Go Pro with an initial learning rate of 1 10 4 for 200K iterations. Then, similar to NCER, we fine-tune the model trained on BSD with the same hyperparameter settings. |