search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.65k stars 7.96k forks source link

Strange results #6502

Closed IDGate-ai closed 4 years ago

IDGate-ai commented 4 years ago

I have trained a detector (yolov3-tiny version with conv.11 weight) to detect 5 objects. The performance is okay, but I want higher accuracy. Then I try yolov4-tiny. They have different result totally. Yolov4-tiny has worse accuracy even the training accuracy (not val acc). And for different objects, yolov3-tiny and yolov4-tiny has different performance. For large object, yolov4-tiny has better accuracy than yolov3-tiny. But for small object, yolov4-tiny has very bad accuracy than yolov3-tiny. Because when I git clone this repository, yolov4 has been released but not release yolov4-tiny. Whether or not @AlexeyAB change something for yolov4-tiny (not include in yolov4)?

IDGate-ai commented 4 years ago

I found out yolov4-tiny-3l will have better result for my dataset. For training and val accuracy with input size is 416*416, yolov4-tiny-3l >> yolov3-tiny > yolov4-tiny ~= yolov3-tiny-3l