tcourat

Hi. I had the same issue. It was caused by a mismatch between the CUDA version used by Pytorch and the native CUDA version in my computer. I got the...

Same issue there.

> When I want to run train.py in the tools directory, it prompts me that mmcv=1.0.5 is required. Does the training of this program only work in this environment? I...

I am interested by the training config for DepthFormer and PackNet too.

> Probably I understood paper wrong, but thought it was mentioned linear classification over features patch-wise. If that is so, 1x1 convolution on unrolled patches is mathematically equivalent to linear...

Google just released a paper very similar to this one (diffusion models for depth prediction). They claim being able to predict **metric depth** if you know camera FOV : https://diffusion-vision.github.io/dmd/

> I had a follow up question about this - the loss function specified in the paper for semantic preservation (equation 9) computes the sum of all the cosine similarity...

Thanks for answering quickly. This is weird because the DMD paper seemed to obtain results coherent with those from ZoeDepth (they have AbsRel=0.091) on SunRGBD.

Hi, you can keep the embeddings during the inference by using forwards hooks. For instance if you want to store the image encoder features while running the mask generator :...

I guess you need to train a new segmentation head on top of the upsampled features from a frozen backbone ?