AugKD: Ingenious Augmentations Empower Knowledge Distillation for Image Super-Resolution

Authors: Yun Zhang, Wei Li, Simiao Li, Hanting Chen, Zhijun Tu, Bingyi Jing, Shaohui Lin, Jie Hu, Wenjia Wang

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive experiments show that the Aug KD significantly outperforms existing state-of-the-art KD methods across a range of SR tasks. Section 4: EXPERIMENTS, 4.2 RESULTS AND COMPARISON, 4.3 ABLATION ANALYSIS, Table 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, Figure 2, 5, 6, 7, 8 all demonstrate empirical studies with data analysis, comparisons, and performance metrics.
Researcher Affiliation Collaboration Authors are affiliated with 'The Hong Kong University of Science and Technology' (academic), 'Huawei Noah s Ark Lab' (industry), 'Southern University of Science and Technology' (academic), and 'East China Normal University' (academic). This represents a mix of academic and industry affiliations.
Pseudocode No The paper describes methods and processes in text and figures, but does not include any clearly labeled pseudocode or algorithm blocks with structured steps.
Open Source Code No The paper states: 'The proposed KD method is implemented using the Basic SR framework (Wang et al., 2022a) and Py Torch 1.10'. While it mentions an open-source framework used, it does not provide specific access information or a direct link to the authors' own implementation code for the Aug KD methodology described in the paper.
Open Datasets Yes The SR models are trained using 800 images from the DIV2K dataset (Timofte et al., 2017) and evaluated on four benchmark datasets: Set5 (Bevilacqua et al., 2012), Set14 (Zeyde et al., 2012), BSD100 (Martin et al., 2001), and Urban100 (Huang et al., 2015). The models are tested on three benchmark datasets: Real SR (Cai et al., 2019), DReal SR (Wei et al., 2020), and OST300 (Wang et al., 2018). We train the 4 scale EDSR models on a much larger dataset (DF2K: DIV2K+Flickr2K (Timofte et al., 2017) with 3450 images).
Dataset Splits Yes The SR models are trained using 800 images from the DIV2K dataset [...] and evaluated on four benchmark datasets: Set5 (Bevilacqua et al., 2012), Set14 (Zeyde et al., 2012), BSD100 (Martin et al., 2001), and Urban100 (Huang et al., 2015). The models are tested on three benchmark datasets: Real SR (Cai et al., 2019), DReal SR (Wei et al., 2020), and OST300 (Wang et al., 2018). We train the 4 scale EDSR models on a much larger dataset (DF2K: DIV2K+Flickr2K (Timofte et al., 2017) with 3450 images).
Hardware Specification Yes The FLOPs and frames per second (FPS) are computed with a 3 256 256 input image on single V100 GPU of 64GB VRAM. [...] training performed on 4 NVIDIA V100 GPUs.
Software Dependencies Yes The proposed KD method is implemented using the Basic SR framework (Wang et al., 2022a) and Py Torch 1.10.
Experiment Setup Yes During training, the input LR images are randomly cropped into 48x48 patches and augmented with random horizontal and vertical flips and rotations. The models are trained using the Adam optimizer (Kingma & Ba, 2014) with β1 = 0.9, β2 = 0.99, and ϵ = 10 8, with a batch size of 16 and a total of 2.5 105 updates. The initial learning rate is set to 10 4 and decays by a factor of 10 every 105 iterations.