Depth-Anything
Depth-Anything copied to clipboard
LiheYoung
Better resolution of depthmap?
Hi this work is great! I have modified it a bit to read and write EXR files in float and also add the depth in the proper Z-buffer channel with OpenImageIO. I have been experimenting with trying to get better resolution by changing the resize values and it works to a degree when I run 777 pixels but after that up to when i fill up my vram at 24GB at 2072 it looks much worse.
transform = Compose([
Resize(
width=2072,
height=2072,
777 on the left and 2072 to the right
Can I do anything else to get the same nice definition but with crisper edges? Do I need to retrain?
Yes, since our models are trained by resizing images to 518 resolution, if you want to use a much larger resolution, you may need to re-train the model with your own data.
Ahh yes I thought so, I probably dont have the resources to train it myself anyway. From the other issues here it seems you will not release the full code used for training? I am working on marrying your code with NVDS to get it even more temporally stable.
And again really awesome work! Congratulations!
Yes, we currently do not have the plan to release the full training code. We may provide models supporting large resolutions in the future, please stay tuned. Thank you!
Yes I saw you mention that in the paper, up to 1000 pixels. Sounds great if it happens, how much time and hardware would you need? And is it not possible to transfer some learning from this model?
Hi this work is great! I have modified it a bit to read and write EXR files in float and also add the depth in the proper Z-buffer channel with OpenImageIO. I have been experimenting with trying to get better resolution by changing the resize values and it works to a degree when I run 777 pixels but after that up to when i fill up my vram at 24GB at 2072 it looks much worse.
transform = Compose([ Resize( width=2072, height=2072,
777 on the left and 2072 to the right
Can I do anything else to get the same nice definition but with crisper edges? Do I need to retrain?
Hi luude, can you elaborate on how to read and write exr files in this model? Been trying to do it but having troubles since I'm practical an amateur in CV.
Hi!
Yes I can elaborate on this a bit when I get back to my computer, I have been travellinhg a few days so I am sorry for the late reply, you can find anything you need here, https://openimageio.readthedocs.io/en/latest/pythonbindings.html . When I get back into this in the coming week I can give you a more detailed explanation.