LiheYoung / Depth-Anything

[CVPR 2024] Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data. Foundation Model for Monocular Depth Estimation
https://depth-anything.github.io
Apache License 2.0
6.49k stars 499 forks source link

serious issues with the new V2 metric depth anything model and others included in the N models! needs immediate attention! #204

Open donrikk opened 3 weeks ago

donrikk commented 3 weeks ago

@LiheYoung hi so i am loving the eork u guys are doing and the current v2 models are amazing the only problem im having is the metric models u just released thre is a big problem i have the map res at 1024 with augmentation on and i process maps and this is the result.

example of n v2 small

i have a hard time believing this was the intended look of the model release and im convinced something must be wrong here. the depth isnt only completely inaccurate but its completely sending foreground objects straight to the background, the maps look more akin to water paintings as of currently. could you possibly give me a bit more insight as to whether or not this was what you guys saw for the model before release or if this is a problem that occurred during upload?

kaelsonofkrypto commented 3 weeks ago

hi im having problems with both the new any v2 n model and the regular any v3 s model, here is a side by side of any s v2 any s v2 something is wrong where the details are lost compared to the hugging face i tested any sv2

LiheYoung commented 3 weeks ago

Hi @donrikk, which model is used?

donrikk commented 3 weeks ago

@LiheYoung the model being used is the "depth_anything_v2_metric_hypersim_vits" but same problems are also present in the base and large model metric versions as well, less in large but still same problems. the kitti models are not really any different from quality as the example above as well so pretty much all of the newest metric v2 models are kind of out of wack i think.

donrikk commented 3 weeks ago

@LiheYoung also some are having issues with v2 small model unrelated to the new metric models. it seems to produce depthmaps that dont match to the v2 small comparison demo online. it seems to be cutting off corners and seems to be haze like compared to how sharp the online comparison generator produces and we are testing the map locally at high res like inthe thousand range so it isnt a qualtiy discrepency between the online hugging face example and local. this seems like a much more minor issue wit hthat map but the newest metric models jsut release seem to be doing something super unnatural lol

loawizard commented 3 weeks ago

thank you Liheyoung for your work! Always have been great fan of anydepth ! Having similar results with depthmap like donrikk. I don't know if thats normal or something went wrong with the depth models? especially new metric maps.

kaelsonofkrypto commented 3 weeks ago

@LiheYoung Hi so we have found the issue, apparently the models look worse at high depth resolution, we found out that using a 518 depth resolution on any s, any n , and any l looks good, but the moment u try to increase beyond 518 depth resolution like 1400 depth resolution, the model looks bad, can you all look into tweaking the models to look better beyond 518 as high depth resolution i believe will fix the gap of making it look like hollywood 3d, here are examples

image

image