Problem when applying this method to one-stage detector about crowddet HOT 7 CLOSED

I have not tried this method on RetinaNet, sorry I can't give you useful suggestions.

from crowddet.

I'll try to implement a one-stage version based on RetinaNet and update the code if it works.
I think how to combine the Emd Loss and Focal Loss is the key to solve the problem.

from crowddet.

Thank you for your response. In fact, I figured out that the drop in ap was due to a bug in my code. However, this method has the problem of giving duplicate predictions to one person compared to the model without this method implemented. The recall is improved for certain precision (+1.4%).
Thank you for your advice. But I do not have much idea about this. I will have a try.

from crowddet.

Applying emd loss to the Retinanet can also have gain in all evaluation indicators.
The result of a simple Retinanet version is as follows, and the code will be released soon.
retina_fpn
mAP:0.8180, mMR:0.5619, JI:0.7312
retina_emd
mAP:0.8303, mMR:0.5537, JI:0.7370

from crowddet.

Thank you for your response. I will wait for your code.

from crowddet.

The new code has included Retinanet implementation.

from crowddet.

The new code has included Retinanet implementation.

Hello, I'm using your Retinanet to do inference,but I found that the inference speed is only a little faster than rcnn, and the inference takes a long time, can you tell me what's going on?

from crowddet.