Use when the H and W are different ...

[lingblessing]

If I have different image H and W, can I still use DPT? What if I want to use pictures with different H and W?

[volgachen]

Yes, you can use it. Different H and W only mean that you will get a different amount of patches. It has no influence how the model processes the image.

I recommend you refer model for detection. The model file here can automatically handle pictures with different sizes.

model = dpt_tiny()
output = model(inputs)

[volgachen]

You can ignore the module argument img_size and patch_count in detection code. It exists just because I also use this code for fixed-resolution classification.

[lingblessing]

How can I ignore img_size and patch_count, these two variables are used for model building.

[volgachen]

Yeah, but you may find that it actually has no influence on the model. I will update the code and remove them later.

[lingblessing]

When will you update? Sorry to urge you, because I have been using DPT these days, and at the same time, I will continue to study how to get rid of the limitations of these two variables. But I still hope you update the code as soon as possible.

[volgachen]

I suggest that you can try set img_size=224 and patch_count=224/patch_size in spite of other contents. It will not influence final results. Update for code may need some days.

[lingblessing]

Okay, I found it, thank you.

[lingblessing]

from .box_coder import *
from .depatch_embed import Simple_DePatch

box_coder = pointwhCoder(*args, **kwargs)
my_module = Simple_DePatch(box_coder, *args, **kwargs)
At this time, is everything in my_module DPT or everything including VIT transformer?
In short, do I need to integrate the DeformablePatchTransformer class?

[volgachen]

from .box_coder import * from .depatch_embed import Simple_DePatch

box_coder = pointwhCoder(*args, **kwargs) my_module = Simple_DePatch(box_coder, *args, **kwargs) At this time, is everything in my_module DPT or everything including VIT transformer? In short, do I need to integrate the DeformablePatchTransformer class?

Yeah, you need to. In this way, my module is only a patch embedding module. If you want to use it with your other backbone model, you need to build the model by yourself

[lingblessing]

May I ask which part of DETR did you embed DPT into? Because the accuracy didn't improve after I added it to DETR, I'm asking you. thanks.