Temporally Consistent Atmospheric Turbulence Mitigation with Neural Representations
Authors: Haoming Cai, Jingxi Chen, Brandon Feng, Weiyun Jiang, Mingyang Xie, Kevin Zhang, Cornelia Fermuller, Yiannis Aloimonos, Ashok Veeraraghavan, Chris Metzler
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive evaluations, we demonstrate that Con VRT substantially improves temporally consistency while also marginally improving per-frame restoration quality. We adopt several real-world datasets for evaluation, including the OTIS (63), Heat Chamber (5), subset of BVI-CLEAR dataset (64), TSR-WGAN dataset (4) and DOST (65). |
| Researcher Affiliation | Academia | Haoming Cai* 1, Jingxi Chen 1, Brandon Y. Feng2, Weiyun Jiang3, Mingyang Xie1, Kevin Zhang1, Cornelia Fermuller1, Yiannis Aloimonos1, Ashok Veeraraghavan3, Christopher A. Metzler 1 1University of Maryland, 2Massachusetts Institute of Technology, 3Rice University |
| Pseudocode | No | The paper describes the pipeline and components in text and diagrams (Figure 3, Figure 4) but does not include formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | More information can be found on our project page: https://convrt-2024.github.io/. We have a website to host all open-source resources, including code, data, and results. The website link is attached to the abstract of the main paper. |
| Open Datasets | Yes | We adopt several real-world datasets for evaluation, including the OTIS (63), Heat Chamber (5), subset of BVI-CLEAR dataset (64), TSR-WGAN dataset (4) and DOST (65). |
| Dataset Splits | No | The paper mentions using several datasets for evaluation and training the model individually on each video clip, but it does not specify explicit training, validation, and test splits with percentages or sample counts for model generalization. |
| Hardware Specification | Yes | Training was conducted on a single RTX A6000. |
| Software Dependencies | No | The paper mentions using the Adam optimizer (66) and Mi Das (60), but it does not provide specific version numbers for any software libraries, frameworks, or programming languages used in the implementation. |
| Experiment Setup | Yes | We trained the Con VRT model individually on each video clip with a learning rate of 2 10 3, using the Adam optimizer (66). For each video clip, the batch size equals to the number of frames in that clip. The temporal resolution parameter Tres was set to 5, with parameters Q1 and Q2 configured to 128 and 256, respectively. |