Unsupervised Flow-Aligned Sequence-to-Sequence Learning for Video Restoration
Authors: Jing Lin, Xiaowan Hu, Yuanhao Cai, Haoqian Wang, Youliang Yan, Xueyi Zou, Yulun Zhang, Luc Van Gool
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that our method achieves state-of-the-art performance in three typical video restoration tasks, including video deblurring, video super-resolution, and compressed video quality enhancement.Datasets. For video SR, the benchmark datasets consist of REDS4 (Nah et al., 2019a) and Vimeo-90K-T (Xue et al., 2019). For video deblurring, we use the GOPRO dataset (Nah et al., 2017), where 22 videos are used for training and 11 videos for testing. For compressed video enhancement, our models are trained with the MFQEv2 dataset (Guan et al., 2019) including 108 lossless videos. We adopt the dataset from ITU-T (Ohm et al., 2012) containing 18 videos for evaluation. |
| Researcher Affiliation | Collaboration | 1Shenzhen International Graduate School, Tsinghua University 2Huawei Noah s Ark Lab 3ETH Zürich. |
| Pseudocode | Yes | Algorithm 1 Unsupervised Distillation Optical Flow Loss |
| Open Source Code | Yes | https://github. com/linjing7/VR-Baseline |
| Open Datasets | Yes | For video SR, the benchmark datasets consist of REDS4 (Nah et al., 2019a) and Vimeo-90K-T (Xue et al., 2019). For video deblurring, we use the GOPRO dataset (Nah et al., 2017), where 22 videos are used for training and 11 videos for testing. For compressed video enhancement, our models are trained with the MFQEv2 dataset (Guan et al., 2019) including 108 lossless videos. We adopt the dataset from ITU-T (Ohm et al., 2012) containing 18 videos for evaluation. |
| Dataset Splits | Yes | For video deblurring, we use the GOPRO dataset (Nah et al., 2017), where 22 videos are used for training and 11 videos for testing. |
| Hardware Specification | Yes | We use Py Torch to implement our models and train them on 8 Tesla V100 GPUs. |
| Software Dependencies | No | The paper mentions 'We use Py Torch to implement our models' but does not specify a version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | During unsupervised optical flow training, the learning rate is set to 1 10 4. And during restoration training, the initial learning rate of the flow estimator and the other modules are set to 5 10 5 and 2 10 4, respectively. |