I’m a master’s student in computer science at the University of Information Technology (VNU-HCM, Vietnam). I’m submitting my first paper to arXiv (cs.CV, cross-list eess.IV), so I need an endorsement.
Short summary: Overfitted codecs such as Cool-chic and C3 clamp the decoder output inside the loop that is differentiated. On bright, low-contrast photos and on screen content, the distortion gradient can become exactly zero. The encoder then only reduces the rate and returns a flat color image, with no error raised. PSNR, SSIM, MS-SSIM and LPIPS score these outputs no higher than the constant image, so no absolute threshold catches them. A straight-through estimator removes the failure at ≤0.1% BD-rate on Kodak, and the bitstream stays identical for the unmodified decoder. The repo includes a small script that recomputes every number in the paper.
If you can endorse for cs.CV and think the work fits, the code is KGU7X3 (enter it at arxiv.org/auth/endorse after logging in to arXiv). I’m also happy to hear any feedback on the paper, including criticism.
Hi Hung — congrats on the paper, and good instinct posting the code and numbers first.
One honest note on the endorsement ask: I can’t give it to you myself. An endorser has to carry standing in cs.CV — recent papers in the category — and I’d just be spending trust I don’t have. That’s exactly what the mechanism is: the endorser vouches that your paper belongs in cs.CV, not that it’s correct, and that costs real reputation. If I’m wrong about that bar, the check is easy — ask any one colleague with an arXiv account in good standing to try it and see what happens.
What I think actually works best: the authors of the codecs you build on. You already cite Cool-chic and C3 — email those authors directly with the repo and the one-line finding (clamped decoder output → exactly-zero distortion gradient → flat image, no error raised). Paper authors endorse follow-on work on their own methods all the time, and public code with a recompute script is exactly the currency they’re looking for. Your advisor or anyone in your department with a cs.CV paper can also do it in two minutes at Log in to arXiv | arXiv e-print repository.
The straight-through estimator fix at ≤0.1% BD-rate with an identical bitstream is a nice result — that’s the kind of thing that makes a reader say yes.
Thanks a lot for the honest explanation and the practical advice, I appreciate it. I’ll keep the post open in case someone with cs.CV standing finds the work fitting.
Update (Oct 1): following the convention of the work this paper builds on (Cool-chic, C3 and COIN all use eess.IV as their primary category), I have moved the submission’s primary category to eess.IV (Image and Video Processing), with a cross-list to cs.CV. Please use the new code below instead of KGU7X3.
arXiv’s rule for eess.IV endorsers: at least 3 papers in eess.AS, eess.IV, eess.SP or eess.SY, submitted between 3 months and 5 years ago. If that is you and the paper fits, I would be very grateful.