Where is Occlusion-aware Cost Volume implemented?

[JunjieLiuSWU]

Hi, Occlusion-aware Cost Volume is proposed in your paper, but I didn't find where it is implemented in the code, could you show me about it?

[fengziyue]

Hi JunjieLiu:

It is implemented at: https://github.com/AutoAILab/DynamicDepth/blob/d674b462cd1a5b2b84062de0b645bcd3f177ab82/dynamicdepth/networks/resnet_encoder.py#L205

[JunjieLiuSWU]

Hi JunjieLiu:

It is implemented at:

https://github.com/AutoAILab/DynamicDepth/blob/d674b462cd1a5b2b84062de0b645bcd3f177ab82/dynamicdepth/networks/resnet_encoder.py#L205

It sames that dynamic object mask is not used here.

[fengziyue]

Thank you for pointing out the confusion. I just updated some variable names and comments.

The dynamic object mask is used in the DOMD module. After it, the occluded areas will be black. So in here we assume pixels of RGB value < 0.15 to be occluded.

Here we project the occlusion mask of the image to each layers of the cost volume.

[JunjieLiuSWU]

`

Thank you for pointing out the confusion. I just updated some variable names and comments.

The dynamic object mask is used in the DOMD module. After it, the occluded areas will be black. So in here we assume pixels of RGB value < 0.15 to be occluded.

Here we project the occlusion mask of the image to each layers of the cost volume.

Thank you for your explanation, but I have another question, I noticed that occluded areas will be set to zeros in the training phase (trainer.py), but It not applied in the inference phase(evaluate_depth.py), that means Occlusion-aware Cost Volume is not used in the inference phase? And why is that?

[fengziyue]

do you mean here?

[JunjieLiuSWU]

do you mean here?

I mean here in the training phase, but there is not the same operation I the inference phase, and dynamic object mask is not used in the inference phase, so how is the DOMD is applied in the inference phase?

[fengziyue]

Seems the evaluate_depth.py is not the correct version, I'm investigating it. You can use the val() function in the trainer.py now, it also evaluates on the test set.

[JunjieLiuSWU]

Seems the evaluate_depth.py is not the correct version, I'm investigating it. You can use the val() function in the trainer.py now, it also evaluates on the test set.

Thank you!