How to use tood detector in mmocr?

[dsp6414]

tood is a super performance detector in mmdet, how to use it in mmocr.

[Mountchicken]

Hi @dsp6414 MMDet's model shares a different architecture from MMOCR, and we are working hard to refactor MMOCR recently to bridge the gap. Soon MMOCR will be able to call the model in MMDet directly.

[jiarenyf]

Hi @dsp6414 MMDet's model shares a different architecture from MMOCR, and we are working hard to refactor MMOCR recently to bridge the gap. Soon MMOCR will be able to call the model in MMDet directly.

When will the refactor ready ? Thanks ...

[Mountchicken]

Hi @jiarenyf This new feature has already been implemented on dev-1.x branch. The document is on the way.

The implementation is simple, and you can refer to MaskRCNN and FCOS for the usage.

[gaotongxiao]

@jiarenyf Some of the dependencies are still under construction, but they will be ready very soon.

[jiarenyf]

Thanks~

[jiarenyf]

@gaotongxiao @Mountchicken What about using backbone in mm segmentation ...

[gaotongxiao]

@jiarenyf We haven't tested it yet, but it's technically feasible as long as the backbone's output matches the head's signature. Try to modify your config:

custom_imports=dict(imports='mmseg.models', allow_failed_imports=False)
model = dict(backbone=dict(type='mmseg.xxx')...)