search

ZwwWayne / K-Net

[NeurIPS2021] Code Release of K-Net: Towards Unified Image Segmentation
Apache License 2.0
467 stars 45 forks source link

how to do postprocessing #10

Closed luohao123 closed 2 years ago

luohao123 commented 2 years ago

the mask seems not accurate, for a same object, it still could be generate multiple masks which is not good, how to postprocessing supress them?

ZwwWayne commented 2 years ago

This is a common issue for all detectors. You can just set a threshold to suppress them. For methods like DETR/K-Net, we usually just output what the network output and it works.

luohao123 commented 2 years ago

@ZwwWayne I found original code outputs fixed 100 objects. I fixed by using a threshold.