Multiple Bounding boxes around same target

yhenon / pytorch-retinanet

Pytorch implementation of RetinaNet object detection.

Apache License 2.0

2.14k stars 664 forks source link

Multiple Bounding boxes around same target #85

Open abdur4373 opened 5 years ago

abdur4373 commented 5 years ago

Hello great work indeed. I am initially visualizing results on different images. There is a problem that i am getting multiple classification targets instead of one. (Image attached). Can you please tell where this can be controlled? I have tried increasing the score from 0.5 to higher but it starts missing detecting some targets. Is there any other way to this. img_screenshot_28 06 2019

JoshVarty commented 5 years ago

I believe non-maximum suppression should prevent/control this.

See: https://github.com/yhenon/pytorch-retinanet/blob/master/lib/nms/pth_nms.py

I'm surprised it hasn't caught it in your case since the boxes are completely overlapping. Maybe the visualization wasn't using it?