search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.73k stars 7.96k forks source link

mAP worser than lightheadrcnn on a small object dataset #5545

Open LewisJin opened 4 years ago

LewisJin commented 4 years ago

Hi, wanna ask how to improve small object detection performance on v4. I am doing object like traffic light detection, I am getting 88% @50%mAP and 50% @%75 mAP on lighthead rcnn while yolov4 got only 78% and 36% respectively.

Any tricks or settings can do to improve such small object datasets?

WongKinYiu commented 4 years ago

For small objects, you have to change anchors setting, or just use larger input resolution for training. (default input size of light head rcnn is "We resize the shorter edge of image to 800 pixels, and restrict the max size of the longer edge to 1200.")

AlexeyAB commented 4 years ago

Try to train yolov4 with these settings

[net]
width=800
height=800
LewisJin commented 4 years ago

@AlexeyAB using 800 it catched up with lightheadrcnn, does there any tricks more to push limitation forward?

AlexeyAB commented 4 years ago

What is the image size in your dataset? Try to train yolov4 with 

[net]
width=1280
height=800