Saumya Saksena

The honest answer is that neither the paper nor this card has a direct number for that. The 32.08 / 33.06 / 37.41 / 32.16 PSNR figures (Outdoor-Rain, RainDrop, Snow100K-S, Snow100K-L) are all for net_g_best, the checkpoint trained on the synthetic composite, evaluated against those same synthetic test sets. The paper's real-world comparison (RealSnow, 1,329 actual photos from Snow100K's real-world split) is qualitative only, Figure 6 in the paper shows side-by-side crops, there's no PSNR/SSIM table for it. That's not an oversight on the authors' part either, RealSnow has no paired clean ground truth, so a full-reference metric like PSNR isn't computable for it the same way it is for the synthetic sets. That absence of a number is itself informative: it means "where does it diverge" can't currently be answered with a metric, only by looking.

If I reason about it qualitatively rather than measuring it: the usual gap for this kind of model shows up where the synthetic renderer's assumptions break down. Snow100K's synthetic particles are fairly uniform in size and motion; real snow varies a lot more in particle size, translucency, and motion blur within a single frame, so I'd expect net_g_real to hold up better there than net_g_best would on the same photo. Outdoor-Rain's fog is a synthetic atmospheric-scattering model applied on top of rain streaks; real rain-plus-fog scenes usually also mix in uneven lighting and haze density that the synthetic pipeline doesn't fully capture, so that's probably where the bigger real-world gap sits too.

That's domain reasoning, not a measured result, and I'd rather say that plainly than imply a number I don't have. If it's useful, I do have both checkpoints and the real vs synthetic test sets locally already (that's what the demo's checkpoint selector is built from), so I could run an actual informal side-by-side and post real examples if there's interest. Happy to do that as a follow-up if it'd help more than the qualitative answer above.

I have the latency and power draw results of each of the model separately. I am planning to make a dedicated article about the benchmarking of the models on a dedicated Drive AGX Orin. As for the cross-camera transfer results, I am not sure I understand completely what you mean.

Hi @Yusufchy
You can find the per-country comparison inside each model card now. Also, please refer to this article here for more details:

https://huggingface.co/blog/dronefreak/gwhd-model-zoo

Hi @Yusufchy
Thank you for your query. You make a great point in the sense that per-country or genotype slicing could indeed impact the results significantly. These stratified results are not yet published in the model cards.

However, if this is something you think could be interesting, I could work upon it and release something like that in the next release. On my side as well I was wondering what could be possible next steps, so thank you for shedding some light in this regard.

I have updated all licenses to AGPL-3 which are the ones used by the original YOLO models.

Thanks for your kind words, and I am glad you found the models useful. You are right to remark about the yolo26-nano performance. I will be starting shortly trying to distill a yolov8s from yolo26-x, in the hopes of further improving the performance.

Also, thanks for your licensing remarks, I will update the licenses of the model cards to AGPL to follow the upstream model licenses.

If you can give me your email id, I can add you to discord or some other form of communication channel.

Aag the fire!

I am glad to know you are interested in this work. I am going to make an initial plan and share it with you ASAP. We can collaborate on GitHub. It will be nice to have some help on this as I myself dont have a lot of time often.

Thanks!