Motion-adaptive Transformer for Event-based Image Deblurring
Authors: Senyan Xu, Zhijing Sun, Mingchen Zhong, Chengzhi Cao, Yidi Liu, Xueyang Fu, Yan Chen
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive testing confirms that our approach sets a new benchmark in image deblurring, surpassing previous methods by up to 0.60d B on the Go Pro dataset, 1.04d B on the HS-ERGB dataset, and achieving an average improvement of 0.52d B across two real-world datasets. |
| Researcher Affiliation | Academia | University of Science and Technology of China EMAIL, EMAIL |
| Pseudocode | No | The paper provides mathematical formulations and descriptions of the architecture components (e.g., Adaptive Motion Mask Predictor, Motion-Sparse Attention) but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://github.com/QUEAHREN/MAT |
| Open Datasets | Yes | Datasets (1) Go Pro (Nah, Hyun Kim, and Mu Lee 2017) is a widely recognized benchmark for motion deblurring. (2) HS-ERGB (Tulyakov et al. 2021) consists of sharp videos and real-world events; (3) REBlur (Sun et al. 2022) collects sequences of real-world events corresponding with real blurry images and sharp images. (4) REVD (Kim, Cho, and Yoon 2024) provides 21 sequences of real-world blursharp image pairs and event streams. |
| Dataset Splits | No | The paper mentions using 'Go Pro and the HS-ERGB datasets' and 'REBlur and the REVD datasets' for testing, but it does not specify explicit train/validation/test split percentages or sample counts for these datasets in the main text. |
| Hardware Specification | Yes | We implement our proposed network via the Py Torch 1.8 platform on NVIDIA RTX 3090 GPU. |
| Software Dependencies | Yes | We implement our proposed network via the Py Torch 1.8 platform on NVIDIA RTX 3090 GPU. Adam W (Loshchilov and Hutter 2017) optimizer with parameters β1 = 0.9 and β2 = 0.99 is adopted to optimize our network. |
| Experiment Setup | Yes | Adam W (Loshchilov and Hutter 2017) optimizer with parameters β1 = 0.9 and β2 = 0.99 is adopted to optimize our network. The learnable expansion factor η and motion modulation factor β are initialized to 0.5 and 1.0, respectively. We train our network with patch size 128 128 and batch size 20 using the charbonnier loss (Charbonnier et al. 1994). The initial learning rate is 5 10 4 and changes with Cosine Annealing scheme to 1 10 6, including 320K iterations in total. Note that we set the bin size of event voxels as 6. |