search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.63k stars 7.95k forks source link

About b-box closely or be looser #7641

Open jayer95 opened 3 years ago

jayer95 commented 3 years ago

@AlexeyAB Hi AlexeyAB,

I’m currently doing model training for license plate characters,

For example, in the following two images, regarding the label bounding box of "1", should I label it closely to "1" or be looser.

Which one can increase the accuracy of training?

ddd

rfffff

In addition, regarding the detection of license plate characters, is there a recommended model architecture and input shape? Can my input shape be very low?

When I use yolov4-tiny for training, I use 128x128 input shape, it accuracy will be reduced to less than half of the original 416x416 (mAP@50: 95%).

stephanecharette commented 3 years ago

You may be interested in this: https://github.com/stephanecharette/DarkPlate

I would label it like the first image.

jayer95 commented 3 years ago

@stephanecharette Thanks for sharing!

What is your reason for choosing to label as the first image? Do you use 37 classes and infer them twice? First detect the license plate and then detect the characters in the license plate ROI?

stephanecharette commented 3 years ago

What is your reason for choosing to label as the first image?

To give it context. So the straight line of 1 doesn't look like the left line in L or the line in R, etc.

Do you use 37 classes and infer them twice? First detect the license plate and then detect the characters in the license plate ROI?

Exactly.