SAHI with latest MMDet/MMRotate/MMYOLO

[hhaAndroid]

Hi, Thank you so much for your wonderful library. I'm an MMDet/MMYOLO developer. We recently made an architecture upgrade and open-sourced MMEngine/MMYOLO. I would like to know if it is possible to collaborate further to support OpenMMLab 2.0 in the sahi repo.
Looking forward to your reply!

[fcakyon]

@hhaAndroid thanks a lot for contacting us!

We are open to such collaboration but I could not understand what changes should be done while updating from mmdet==2.25.1 mmcv-full==1.6.1 to OpenMMLab 2.0. Would you mind guiding us?

[hhaAndroid]

We will provide a draft design. Do you have any contact information there?

[fcakyon]

You can mail me at fatihcagatayakyon[at]gmail[dot]com

[hhaAndroid]

OK

[fcakyon]

Furthermore, SAHI supports instance segmentation outputs, mmrotate support can be added to SAHI by handling mmrotate outputs as 4 points polygons.

[hhaAndroid]

@fcakyon Hi, we will start this project next month. Before we do that, I would like to ask if you have any technical documentation available? Or about the support process for mmdet in sahi?

For mmdet 3.x and mmyolo you can get some information from https://mmdetection.readthedocs.io/en/3.x/ and https://mmyolo.readthedocs.io/en/latest/. We will also be releasing an English PR release soon.

[fcakyon]

Hello @hhaAndroid, we use high level api for mmdet integration.

init_detector:

sahi/sahi/models/mmdet.py

Line 28 in 2ddf8a8

model = init_detector(

inference_detector:

sahi/sahi/models/mmdet.py

Line 78 in 2ddf8a8

prediction_result = inference_detector(self.model, image)

are there any changes in the usage of these init_detector and inference_detector functions?

[hhaAndroid]

The change is very small. I will try it in MMYOLO. I'll let you know once it's done

[thiagoribeirodamotta]

Was sahi ever added for training on MMDet?

[Rosal-1998]

i wonder can we use sahi in mmyolo now? thank you for your reply.