search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.71k stars 7.96k forks source link

NMS duplicate bounding box - tiny yolo3 #6459

Open TTamyy opened 4 years ago

TTamyy commented 4 years ago

NMS duplicate bounding box - tiny yolo3

Hi @AlexeyAB i trained on dataset with multiple classes, but as class bicycle and class motorbike is same bounding box - xywh is same but different classes, with same same conf score.

My Question:

  1. Does nms thres only works with same class but ignore other classes?
  2. Is there anything in side darknet can change to make the prediction better and exclude cases like this?
AlexeyAB commented 4 years ago

  1. Yes.

  2. If motorbike is detected as motorbike and bicycle, then there is just false-detection (bicycle). There is no special way to remove false-detection, except to increase the accuracy of the detector - try to use: https://github.com/AlexeyAB/darknet#how-to-improve-object-detection You can try to increase -thresh 0.3 or 0.5 it will decrease false-positives, but it will increase false-negatives too.