Controllable Blur Data Augmentation Using 3D-Aware Motion Estimation
Authors: Insoo Kim, Hana Lee, Hyong-Euk Lee, Jinwoo Shin
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To demonstrate the effectiveness of our blur synthesizer, we conduct experiments across various datasets such as Go Pro (Nah et al., 2017), HIDE (Shen et al., 2019), Real Blur-J and R (Rim et al., 2020), and network architectures such as MIMO-UNet+ (Cho et al., 2021), Restormer (Zamir et al., 2022), NAFNet (Chen et al., 2022) and FFTformer (Kong et al., 2023). We train with Go Pro and evaluate Go Pro and HIDE. Also, we train with Real Blur-J and R, and evaluate the corresponding trained models. As shown in Table 1, the results clearly demonstrate that our blur synthesizer, i.e., Geo Syn boosts deblurring performance across not only different network architectures but also various datasets. |
| Researcher Affiliation | Collaboration | Insoo Kim1,2 Hana Lee1 Hyong-Euk Lee1 Jinwoo Shin2 1SAIT, Samsung Electronics 2KAIST |
| Pseudocode | Yes | The pseudo-codes can be found in Section G and H of Appendix. G VECTOR FIELD GENERATION IN PYTORCH PSEUDOCODE. H BLUR SYNTHESIS AND DATA AUGMENTATION IN PYTORCH PSEUDOCODE. |
| Open Source Code | No | REPRODUCIBILITY STATEMENT To facilitate reproducibility, we present comprehensive implementation details for training our blur synthesizer in Section 4.1 and Fig. 2. We also describe the details of the data augmentation strategy, i.e., where to synthesize, how to synthesize, and how to augment during the training of the deblurring model, in Section B of Appendix. The hyperparameter settings are empirically optimized as described in Section 4.5 and Section F.1 of Appendix, and are summarized in Section 4.1. We also provide the pytorch pseudo-codes of our blur synthesis and blur data augmentation in Section G and H of Appendix. |
| Open Datasets | Yes | For datasets, we use Real Blur (Rim et al., 2020), RSBlur (Rim et al., 2022), and Go Pro (Nah et al., 2017) datasets for training and evaluation. |
| Dataset Splits | Yes | Each Real Blur type consists of 3,758 and 980 image pairs for training and test sets, respectively. RSBlur contains 8,878 and 3,360 blur-sharp image pairs for training and test sets, respectively. Go Pro contains 2,103 and 1,111 image pairs for training and test sets, respectively. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU models, or memory amounts used for running experiments. It mentions NAFNet model sizes (e.g., NAFNet-16, NAFNet-32, NAFNet-64) but these refer to the neural network architecture, not the hardware. |
| Software Dependencies | No | The paper mentions using the Adam W optimizer and PyTorch (implicitly, through pseudocode), but does not specify version numbers for any software components (e.g., Python, PyTorch, CUDA). |
| Experiment Setup | Yes | We train our blur synthesizer up to 1, 000 epochs for Real Blur, 3, 000 epochs for Go Pro, and 500 epochs for RSBlur. Also, our blur synthesizer is optimized by the Adam W (Loshchilov & Hutter, 2019) algorithm (β1 = 0.9, β2 = 0.9 and weight decay 1e 3) with the cosine annealing schedule (1e 3 to 1e 7) (Loshchilov & Hutter, 2016) gradually reduced for total iterations of each dataset. Unless otherwise specified, we use λ1 = 0.1, λ2 = 1.0 and the number of camera positions as 16, which is discussed in Section 4.5 and F.1 of Appendix. We use Real Blur-J for all ablation studies. The pseudo-codes can be found in Section G and H of Appendix. |