JPEGAI performance on PSNR-RGB
I’m currently evaluating JPEG-AI on PSNR-RGB. In particular, I’m focusing on the SOP decoder (without tools) because of its low decoding complexity (~8000 MAC).
I’m using the main branch of the repository, which I believe includes the latest models. Here’s the command I run:
python -m src.reco.scripts.eval --coding_type enc_dec --in_dir /path/to/dataset --out_dir results/sop --cfg cfg/tools_off.json cfg/profiles/simple.json
I realize JPEG-AI is not specifically optimized for this evaluation setup, but I’m observing some unexpected results. On the CLIC dataset, I measure:
- +14% BD-Rate vs HM with HOP
- +38% BD-Rate vs HM with SOP
The RD curve also looks unusual, especially at the high-rate point:
Do you know if this behavior is expected? Are there any common pitfalls or configuration issues that could explain such results?
Thanks a lot, Théophile Blard