Learning to Inpaint for Image Compression
Authors: Mohammad Haris Baig, Vladlen Koltun, Lorenzo Torresani
NeurIPS 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate performance, we perform compression with our models on images from the Kodak dataset [8]. The dataset consists of 24 uncompressed color images of size 512 768. The quality is measured according to the MSSSIM [22] metric (higher values indicate better quality). We use the Bjontegaard-Delta metric [4] to compute the average reduction in bit-rate across all quality settings. |
| Researcher Affiliation | Collaboration | Mohammad Haris Baig Department of Computer Science Dartmouth College Hanover, NH Vladlen Koltun Intel Labs Santa Clara, CA Lorenzo Torresani Dartmouth College Hanover, NH |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. It describes methods in paragraph form and through architectural diagrams. |
| Open Source Code | No | The paper provides a URL (http://www.cs.dartmouth.edu/~haris/compression) which appears to be a personal project page, not an explicit statement of code release or a direct link to a code repository for the methodology described in the paper. |
| Open Datasets | Yes | Our models were trained on 6,507 images from the Image Net dataset [7], as proposed by Ballé et al. [2] to train their single-stage encoder-decoder architectures. |
| Dataset Splits | No | The paper does not provide specific dataset split information (exact percentages, sample counts, or citations to predefined splits) for training, validation, and testing needed to reproduce the data partitioning. It mentions the number of images in the training dataset and the dataset used for evaluation but not the splits. |
| Hardware Specification | No | The paper states "We gratefully acknowledge NVIDIA and Facebook for the donation of GPUs used for portions of this work," but it does not provide specific hardware details such as exact GPU/CPU models or processor types. |
| Software Dependencies | No | The paper states "We use the Caffe library [11] to train our models," but it does not specify the version number of Caffe or any other software dependencies. |
| Experiment Setup | Yes | The residual encoder and R2I models were trained for 60,000 iterations whereas the joint inpainting network was trained for 110,000 iterations. We used the Adam optimizer [12] for training our models and the MSRA initialization [9] for initializing all stages. We used initial learning rates of 0.001 and the learning rate was dropped after 30K and 45K for the R2I models. For the IR2I model, the learning rate was dropped after 30K, 65K, and 90K iterations by a factor of 10 each time. All of our models were trained to reproduce the content of 32 32 image patches. |