Improving Inference for Neural Image Compression

Authors: Yibo Yang, Robert Bamler, Stephan Mandt

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, which include extensive baseline comparisons and ablation studies, we achieve new state-of-the-art performance on lossy image compression using an established VAE architecture, by changing only the inference method.
Researcher Affiliation Academia Yibo Yang, Robert Bamler, Stephan Mandt Department of Computer Science University of California, Irvine {yibo.yang, rbamler, mandt}@uci.edu
Pseudocode Yes Algorithm 1: Proposed lossy bits-back coding method (Section 3.3 and Figure 1e-f).
Open Source Code No The paper does not provide any statement or link for the open-source code of the described methodology.
Open Datasets Yes We improve its performance drastically, achieving an average of over 15% BD rate savings on Kodak and 20% on Tecnick [Asuni and Giachetti, 2014]Eastman Kodak. Kodak lossless true color image suite (Photo CD PCD0992). URL http://r0k. us/graphics/kodak.N. Asuni and A. Giachetti. TESTIMAGES: A large-scale archive for testing visual devices and basic image processing algorithms (SAMPLING 1200 RGB set). In STAG: Smart Tools and Apps for Graphics, 2014. URL https://sourceforge.net/projects/testimages/files/OLD/ OLD_SAMPLING/testimages.zip.
Dataset Splits No The paper uses the Kodak and Tecnick datasets for experiments but does not provide specific training, validation, or test dataset splits.
Hardware Specification No The paper does not provide specific hardware details (GPU/CPU models, memory, etc.) used for running its experiments.
Software Dependencies Yes In all results, we used Adam [Kingma and Ba, 2015] for optimization, and annealed the temperature of SGA by an exponential decay schedule, and found good convergence without per-model hyperparameter tuning.
Experiment Setup Yes In all results, we used Adam [Kingma and Ba, 2015] for optimization, and annealed the temperature of SGA by an exponential decay schedule, and found good convergence without per-model hyperparameter tuning. We provide details in the Supplementary Material.