Real-Time Adaptive Image Compression

Authors: Oren Rippel, Lubomir Bourdev

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our algorithm outperforms all existing image compression approaches, both traditional and ML-based: it typically produces files 2.5 times smaller than JPEG and JPEG 2000 (JP2), 2 times smaller than Web P, and 1.7 times smaller than BPG on the Kodak Photo CD and RAISE-1k 512 768 datasets across all of quality levels. At the same time, we designed our approach to be lightweight and efficiently deployable. On a GTX 980 Ti GPU, it takes around 9ms to encode and 10ms to decode an image from these datasets: for JPEG, encode/decode times are 18ms/12ms, for JP2 350ms/80ms and for Web P 70ms/80ms. Results for a representative quality level are presented in Table 1.
Researcher Affiliation Industry 1Wave One Inc., Mountain View, CA, USA.
Pseudocode No No structured pseudocode or algorithm blocks found. The paper describes the model architecture and components in descriptive text and figures.
Open Source Code No No statement explicitly providing concrete access to the source code for the methodology described in this paper was found. The paper mentions other authors' code availability but not its own.
Open Datasets Yes We trained all models on 128 128 patches sampled at random from the Yahoo Flickr Creative Com-mons 100 Million dataset (Thomee et al., 2016).
Dataset Splits No No specific dataset split information (e.g., percentages, counts, or predefined splits) for training, validation, and testing of their *own* model on the Yahoo Flickr Creative Commons dataset is provided. The paper discusses test sets (Kodak, RAISE-1k) separately from the training data source.
Hardware Specification Yes We trained and tested all models on a Ge Force GTX 980 Ti GPU and a custom codebase.
Software Dependencies No No specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment were found. The paper mentions optimizing with 'Adam' but does not provide software versions for frameworks or libraries.
Experiment Setup Yes We trained all models on 128 128 patches sampled at random from the Yahoo Flickr Creative Com-mons 100 Million dataset (Thomee et al., 2016). We optimized all models with Adam (Kingma & Ba, 2014). We used an initial learning rate of 3 10 4, and reduced it twice by a factor of 5 during training. We chose a batch size of 16 and trained each model for a total of 400,000 iterations. We initialized the ACR coefficient as α0 = 1.