Awesome

Region Adaptive Transform with Segmentation Prior for Image Compression

The paper has been accepted by ECCV2024! Thank you for your attention!

About

Our SegPIC introduces proposed RAT and SAL based on WACNN.

We compare our SegPIC with previously well-performing methods.

Visualization of the reconstructed images kodim04 and kodim24 in Kodak. The metrics are (PNSR↑/bpp↓). It shows that our SegPIC can distinguish the objects’ contours more accurately, making the edges sharper with less bitrate.

Installation

The code is based on WACNN and CompressAI. You can refer to them for installation. It is also recommended to adopt Pytorch-2.0 for faster training speed.

Checkpoints

We provide 6 checkpoints optimized by MSE. See Google Drive.

Training Dataset

COCO-train-2017 for training, COCO-val-2017 for validation and panoptic_annotations for .png masks. Images and masks correspond by the same filename (no suffix). The data format is as follows:

- COCO-Stuff/
    - train2017/
        - img000.jpg
        - img001.jpg
    - val2017/
        - img002.jpg
        - img003.jpg
    - annotations/
        - panoptic_train2017/
            - img000.png
            - img001.png
        - panoptic_val2017/
            - img002.png
            - img003.png

Training and Testing

The overall usage is the same as WACNN and CompressAI. Please see run.sh and test.sh.