search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.68k stars 7.96k forks source link

Exclude crowd detections #4129

Open csestili opened 4 years ago

csestili commented 4 years ago

Hi, thank you so much for providing a usable YOLOv3 batch processor!

When a crowd of people are present in an image, sometimes the model will detect them all with one person bounding box. See this example: image

This is expected behavior, because the COCO dataset contains person bounding boxes that encompass crowds, with a special metadata value iscrowd = 1. Is there a way to get the model not to output bounding boxes for crowds? E.g., does it predict the iscrowd label as well, so that I can filter crowd boxes out?

AlexeyAB commented 4 years ago

It seems this is related to this issue: https://github.com/AlexeyAB/darknet/issues/4085 So you can try to improve MS COCO dataset.

Also, even if training dataset is correct, then NMS can lead to fusion many bboxes of the same class_id to one bounded boxes.