search

lkeab / BCNet

Deep Occlusion-Aware Instance Segmentation with Overlapping BiLayers [CVPR 2021]
https://arxiv.org/abs/2103.12340
MIT License
529 stars 74 forks source link

if the Occluder is not in the category tag #110

Closed wmrenr closed 2 years ago

wmrenr commented 2 years ago

If the Occluder marked in the json file is not in the category tags,can it be identified？

lkeab commented 2 years ago

During training, if it is class-agnostic for occluder mask prediction, then it should be able to identify.

wmrenr commented 2 years ago

what's the mean of 'it is class-agnostic for occluder mask prediction' ? Can you explain in detail

wmrenr commented 2 years ago

@lkeab
The pictures of my dataset are tomatoes that are occluded by leaves to varying degrees. The categories of the dataset are tomatoes and leaves. I have annotated both tomatoes and leaves, but the category id in each of my annotation files is set to tomatoes. The experimental result is that I can visualize the tomatoes but not the leaves. Can you help me ?

lkeab commented 2 years ago

can you set the training into class-agnostic? That means the mask head will only predict the target object mask with single channel dimension instead of C (class number) dimensions.

wmrenr commented 2 years ago

How to set the training into class-agnostic ? By by modifying the codes ?

wmrenr commented 2 years ago

@lkeab I found class-agnostic detector and class-aware detector on the Internet, for this problem, I need to change the detector?

lkeab commented 2 years ago

just change the mask head setting. It's not necessary to change the whole detector.

wmrenr commented 2 years ago

just change the mask head setting. It's not necessary to change the whole detector. In which code should I change the mask head setting？Can you give some hints？

lkeab commented 2 years ago

You can set it here

wmrenr commented 2 years ago

You can set it here Ok,thanks.