EvTexture: Event-driven Texture Enhancement for Video Super-Resolution
Authors: Dachun Kai, Jiayao Lu, Yueyi Zhang, Xiaoyan Sun
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that our Ev Texture achieves state-of-the-art performance on four datasets. |
| Researcher Affiliation | Collaboration | 1University of Science and Technology of China 2Institute of Artificial Intelligence, Hefei Comprehensive National Science Center. |
| Pseudocode | No | The paper describes its methods through textual explanations and network diagrams (e.g., Figure 3) but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks with structured steps. |
| Open Source Code | Yes | Code: https: //github.com/Dachun Kai/Ev Texture. |
| Open Datasets | Yes | Synthetic datasets. We use two popular datasets for training: Vimeo-90K (Xue et al., 2019) and REDS (Nah et al., 2019). Real-world datasets. Following previous event-based VSR studies (Lu et al., 2023; Kai et al., 2023), we use the CED (Scheerlinck et al., 2019) dataset for training and evaluating on real-world scenes. |
| Dataset Splits | Yes | Specifically, for Vimeo-90K, Vid4 (Liu & Sun, 2013) and Vimeo-90K-T serve as our test sets... For REDS, we employ REDS4... as our test set and evaluate in the RGB channel. ...we use the CED (Scheerlinck et al., 2019) dataset for training and evaluating on real-world scenes... we select 11 clips2 from the total of 84 clips as our test set and use the remainder for training. Vimeo-90K... The dataset includes 64,612 training clips and 7,824 testing clips, known as Vimeo-90K-T. We categorize these clips into three difficulty levels: easy, medium, and hard... we classify the first 50% (3,907 clips) as easy, the next 30% (2,345 clips) as medium, and the final 20% (1,563 clips) as hard. |
| Hardware Specification | Yes | The whole training is conducted on 8 NVIDIA RTX3090 GPUs and takes about four days. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer (Kingma & Ba, 2015)', 'Cosine Annealing (Loshchilov & Hutter, 2016) scheduler', 'Charbonnier penalty loss (Lai et al., 2017)', and 'pre-trained model Spy Net (Ranjan & Black, 2017)'. However, it does not specify software versions for these tools or for the programming environment (e.g., Python version, PyTorch version, CUDA version). |
| Experiment Setup | Yes | We use 15 frames as inputs for training and set the minibatch size as 8 and the input frame size as 64 64. We augment the training data with random horizontal and vertical flips. We train the model for 300K iterations and adopt Adam optimizer (Kingma & Ba, 2015) and Cosine Annealing (Loshchilov & Hutter, 2016) scheduler. The Charbonnier penalty loss (Lai et al., 2017) is applied for supervision... For Spy Net, the initial learning rate is 2.5 10 5, frozen for the first 5K iterations. The initial learning rate for other modules is 2 10 4. |