Few-Shot Domain Adaptation for Learned Image Compression

Authors: Tianyu Zhang, Haotian Zhang, Yuqi Li, Li Li, Dong Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments across multiple domains and multiple representative LIC schemes demonstrate that our method significantly enhances pre-trained models, achieving comparable performance to H.266/VVC intra coding with merely 25 target-domain samples. Additionally, our method matches the performance of full-model finetune while transmitting fewer than 2% of the parameters. ... We conduct ablation studies to evaluate the contributions of Conv-Adapters, Lo RA-Adapters and the two-stage training strategy.
Researcher Affiliation Academia University of Science and Technology of China EMAIL, EMAIL
Pseudocode No The paper describes the methods in text and uses figures to illustrate deployment and division (Figure 4 and Figure 5), but does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code No The paper does not contain any explicit statements about code availability, such as a direct link to a repository or a declaration that the code will be released or is available in supplementary materials.
Open Datasets Yes We follow Lv et al. (2023); Shen, Yue, and Yang (2023) and sort five domains with 150 images each, including (i) pixel-style images (Lv et al. 2023); (ii) screen content images from SCID (Ni et al. 2017), SCI1K (Yang et al. 2021), SIQAD (Yang, Fang, and Lin 2015); (iii) craters images cropped from Lunar Reconnaissance Orbiter Camera (LROC) (Robinson et al. 2010); (iv) game images cropped from Gaming Video SET (Barman et al. 2018); (v) pathology images cropped from BRACS (Brancati et al. 2022). All crops have the same resolution of 600 800. We select a fixed test set with 100 images for each domain. We use Kodak (Franzen 1993) for the natural image domain.
Dataset Splits Yes We train adapters with N = {5, 10, 25} samples, respectively. We split the samples into training and validation set in a proportion of 4:1. We select a fixed test set with 100 images for each domain.
Hardware Specification No The Acknowledgments section mentions a 'GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC' but does not specify any particular GPU models, CPU details, or other specific hardware components used for running experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers, such as programming languages, libraries, or frameworks used for implementation.
Experiment Setup Yes We train adapters with N = {5, 10, 25} samples, respectively. We split the samples into training and validation set in a proportion of 4:1. We use a patch size of 256 and a batch size of 4. In the first stage, we set learning rate stages {50, 10, 7.5, 5, 2.5, 1} 10 5. For different N, we further set max epoch = {250, 500, 750}. In the second stage, we train selected adapters for another max epoch with fixed learning rate 5 10 4.