Another Way to the Top: Exploit Contextual Clustering in Learned Image Coding

Authors: Yichi Zhang, Zhihao Duan, Ming Lu, Dandan Ding, Fengqing Zhu, Zhan Ma

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the superior performance of CLIC over state-of-the-art works: when optimized using MSE, it outperforms VVC by about 10% BD-Rate in three widely-used benchmark datasets; when optimized using MSSSIM, it saves more than 50% BD-Rate over VVC. Our CLIC offers a new way to generate compact representations for image compression, which also provides a novel direction along the line of LIC development. The paper includes sections like 'Experimental Results', 'Quantitative Results', 'Qualitative Visualization', and 'Ablation Study' with detailed performance tables and figures.
Researcher Affiliation Academia 1Hangzhou Normal University, Hangzhou, Zhejiang, China 2Purdue University, West Lafayette, Indiana, U.S. 3 Nanjing University, Nanjing, Jiangsu, China
Pseudocode No The paper describes algorithmic steps for clustering in prose (steps 1-5 in the 'Contextual Clustering' section), but does not present them in a structured pseudocode or algorithm block.
Open Source Code No The paper does not provide an explicit statement or link for open-source code availability for the described methodology.
Open Datasets Yes We use Flicker2W (Liu et al. 2020b) and LIU4K (Liu et al. 2020a) as our training sets. The paper also provides links to the testing datasets: Kodak2 (https://r0k.us/graphics/kodak/), Tecnick3 (https://tecnick.com/?aiocp%20dp=testimages), and CLIC 20224 (http://compression.cc/).
Dataset Splits Yes Following Liu et al. (Liu et al. 2020b), 99% of the images were used for training, and the remaining 1% were used for validation.
Hardware Specification Yes All training was performed on a computer with an RTX4090 GPU, i9-13900K CPU, and 64G RAM. Ablation experiments were performed on a computer with an RTX3090 GPU, i7-9700K CPU, and 64G RAM.
Software Dependencies No The paper mentions specific optimizers (Adam) and methods (U-Q, DS-Q) with citations, but does not specify versions of programming languages, libraries (e.g., PyTorch, TensorFlow), or other software dependencies.
Experiment Setup Yes Following the settings of Compress AI (B egaint et al. 2020), we set λ {18, 35, 67, 130, 250, 483} 10 4 for MSE optimized model and λ {2.40, 4.58, 8.73, 16.64, 31.73, 60.50} for MS-SSIM optimized model. We trained each model with Adam optimizer (Kingma and Ba 2014) with β1 = 0.9, β2 = 0.999. Each model was trained for 300 epochs with a batch size of 8 and an initial learning rate of 1e 4. We used the Reduce LRon Plateau lr scheduler with a patience of 5 and a factor of 0.5. The first 150 epochs are called Stage 1 and the next 150 epochs are called Stage 2. When switching from Stage 1 to Stage 2, the learning rate is newly set to 1e 4.