Adaptive Window Pruning for Efficient Local Motion Deblurring

Authors: Haoying Li, Jixin Zhao, Shangchen Zhou, Huajun Feng, Chongyi Li, Chen Change Loy

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our proposed LMD-Vi T and baseline methods on the Re Lo Blur testing dataset (Li et al., 2023) with the full image size of 2152 × 1436 on 1 Nvidia A100 GPU. In testing, the inputs are solely of locally blurred images without accompanying blur masks. In addition to the commonlyused PSNR and SSIM (Wang et al., 2004) metrics, we calculate weighted PSNR and weighted SSIM (Li et al., 2023) to better assess the local deblurring performance. We provide the evaluation results in the following sections and Appendix F. Evaluations on public datasets. We first compare the proposed LMD-Vi T with both CNN-based methods (Nah et al., 2017; Kupyn et al., 2019; Chen et al., 2021; Cho et al., 2021; Li et al., 2023) and Transformer-based methods (Zamir et al., 2022; Wang et al., 2022) on the Re Lo Blur dataset (Li et al., 2023) for local motion deblurring. As depicted in Figure 1 and Figure 3, LMD-Vi T exhibits superior performance compared to other state-of-the-art methods, producing clearer outputs with enhanced details. User study on real-world photos. To validate the effectiveness of our proposed model in real-world locally blurred images, we employ a static Sony industrial camera and a static Fuji XT20 SLR camera to capture 18 locally motion-blurred images, each with a resolution of 6000 × 4000 pixels. Subsequently, we conduct a comparative evaluation against the top 3 methods as listed in Table 1. As ground truths for these blurred images are unavailable, we organized a user study involving 30 participants with a keen interest in photography. Analyses. To examine the effectiveness of our proposed method, we conduct separate analyses focusing on the window pruning strategy and blur mask annotations in this section.
Researcher Affiliation Academia Haoying Li1,2 , Jixin Zhao1, Shangchen Zhou1, Huajun Feng2, Chongyi Li3 & Chen Change Loy1 1S-Lab, Nanyang Technological University, Singapore 2Zhejiang University, China 3Nankai University, China
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Codes and data are available at https://leiali.github.io/LMD-Vi T_webpage.
Open Datasets Yes LMD-Vi T is trained on the Go Pro dataset (Nah et al., 2017) and Re Lo Blur dataset (Li et al., 2023) together, in order to enable both local and global motion deblurring.
Dataset Splits No The paper mentions 'The sampling ratio of the Go Pro training data (Nah et al., 2017) and the Re Lo Blur training data (Li et al., 2023) is close to 1:1.' but does not specify a training/validation/test split or specific percentages/counts for a validation set.
Hardware Specification Yes We evaluate our proposed LMD-Vi T and baseline methods on the Re Lo Blur testing dataset (Li et al., 2023) with the full image size of 2152 x 1436 on 1 Nvidia A100 GPU.
Software Dependencies No The paper mentions using 'Adam W optimizer (Kingma & Ba, 2015)' but does not provide specific version numbers for any software dependencies like programming languages (e.g., Python), deep learning frameworks (e.g., PyTorch, TensorFlow), or other libraries (e.g., CUDA).
Experiment Setup Yes We train LMD-Vi T using Adam W optimizer (Kingma & Ba, 2015) with the momentum terms of (0.9, 0.999), a batch size of 12, and an initial learning rate of 2 x 10^-4 that is updated every 2k steps by a cosine annealing schedule (Loshchilov & Hutter, 2017). We set the window size of Ada WPT to 8 x 8, and the initial embedded dim to 32 which is doubled after passing each down-sampling layer.