search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.75k stars 7.96k forks source link

duplicate detection of an object #5269

Open RadouaneK opened 4 years ago

RadouaneK commented 4 years ago

Hi @AlexeyAB and anyone who can help me I just want some explanation for my problem. following the instructions in this repository, I trained darknet model with my custom dataset for detection, but during inference, I sometimes was getting duplicate detection of a particular object i.e multiple bounding boxes. After digging more in this issue, I realized that It happens because I set 'angle' to 90 in the cfg file. I set it back to 0 and retrain again and everything works fine. My question here is: why it happened that way? my thinking is rotation is just another data augmentation technique, so why rotating the images during training will raise this issue?

I hope my question is clear to everyone and I appreciate your answers. thanks

AlexeyAB commented 4 years ago

Currently 'angle=' is allowd only for Classifier, not for Detector.

I sometimes was getting duplicate detection of a particular object i.e multiple bounding boxes.

This is just false-positive detection, because there are no detectors with 100% accuracy