Selective compression learning of latent representations for variable-rate image compression
Authors: Jooyoung Lee, Seyoon Jeong, Munchurl Kim
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The extensive experimental results show that the proposed method can achieve comparable compression efficiency as those of the separately trained reference compression models and can reduce decoding time owing to the selective compression. We train the proposed SCR model in an end-to-end manner, where the base compression model, 3D importance map generation module, γq vectors, QVq vectors, and IQVq vectors are optimized all together, using the total loss formulated as follows: We measure average bpp-PSNR and bpp-MS-SSIM values over the Kodak image set [28]. |
| Researcher Affiliation | Academia | Jooyoung Lee1,2, Seyoon Jeong2, Munchurl Kim1 1School of Electrical Engineering, Korea Advanced Institute of Science and Technology, Korea 2Electronics and Telecommunications Research Institute, Korea |
| Pseudocode | No | The paper describes processes and operators, but does not include structured pseudocode or algorithm blocks labeled as such. |
| Open Source Code | Yes | The sample codes are publicly available at https://github.com/JooyoungLee-ETRI/SCR. |
| Open Datasets | Yes | We measure average bpp-PSNR and bpp-MS-SSIM values over the Kodak image set [28]. [28] E. Kodak, Kodak lossless true color image suite (photocd pcd0992), 1993. [Online]. Available: http://r0k.us/graphics/kodak/ |
| Dataset Splits | No | The paper mentions training and testing on the Kodak image set, but does not provide specific train/validation/test dataset splits or methodologies needed for reproduction. |
| Hardware Specification | Yes | We measured the average decoding time of each model over the Kodak image dataset [28] with 2 Intel Xeon Gold 5122 CPUs and 1 RTX Titan GPU. |
| Software Dependencies | No | The paper mentions using different entropy coders and models (e.g., Gaussian distribution), but does not specify any software names with version numbers for replication. |
| Experiment Setup | Yes | We train the proposed SCR model in an end-to-end manner, where the base compression model, 3D importance map generation module, γq vectors, QVq vectors, and IQVq vectors are optimized all together, using the total loss formulated as follows: L = P q Rq + λq Dq, with Rq = Hq( ys q | z) + H( z), (4) where Rq and Dq represent rate and distortion terms, respectively, for the target quality level q, and λq = 0.2 2q 8 is a parameter for adjusting the balance between the rate and distortion. In our implementation, NQ is set to 8 and Cy is set to that of the original reference model. |