Detectron2: OHEM implementation
Will this project share the OHEM implementation?
WeihongM
All 2 comments
I second the request. OHEM is useful for real-world applications where high precision is a must. These guys have OHEM and a few other useful features: https://github.com/open-mmlab/mmdetection Hope Detectron can support some of these.
I came across mmlab thanks to this https://github.com/microsoft/RepPoints This coule be useful too, many objects of interest are not rectangular / rotated rectangular or even polygonal.
cbasavaraj
on 22 Nov 2019
Having OHEM is very useful for real world tasks in cases where we train specific lightweight backbone net from scratch for specific application. The obvious alternative is to use specific loss, like Focal Loss. From practice (I don't have concrete numbers though) it depends, in some cases one approach gives better results, in some cases another.
We have several practical cases trained with OHEM.
However you don't really need OHEM with a good pretrained backbone, and it's a question if detectron2 considers case of training from scratch for specific use-cases.
ArutyunovG
on 27 Nov 2019
Related issues
ChungNPH
路
3Comments
Cold-Winter
路
3Comments
marcoippolito
路
4Comments
joeythegod
路
4Comments
invisprints
路
4Comments
Most helpful comment
Having OHEM is very useful for real world tasks in cases where we train specific lightweight backbone net from scratch for specific application. The obvious alternative is to use specific loss, like Focal Loss. From practice (I don't have concrete numbers though) it depends, in some cases one approach gives better results, in some cases another.
We have several practical cases trained with OHEM.
However you don't really need OHEM with a good pretrained backbone, and it's a question if detectron2 considers case of training from scratch for specific use-cases.