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

Bounding box shifted towards left or right - Elongated bounding box #8029

Open AkkiSony opened 3 years ago

AkkiSony commented 3 years ago

I trained a custom model with 5 classes. During my inference, I get detections where the bounding box is either shifted towards the left or right. Please have a look at the pictures.

I stopped the training at mAP = 96.4% with an avergae loss of about 0.6, yet the score value of the predictions are very small.

coral-output VIM3-ROTATED-270-output

stephanecharette commented 3 years ago

My first guess is your annotations are not correct.

stephanecharette commented 3 years ago

Have you looked at your annotations using one of the usual tools to ensure they are done correctly?

AkkiSony commented 3 years ago

Have you looked at your annotations using one of the usual tools to ensure they are done correctly?

I myself did the annotations using LabelImg tool

AkkiSony commented 3 years ago

Just another question, which I am curious to know about. The input image into the darknet model is vertical (imagine the VIM3 or google dev board vertical), but when I get the output image, it is rotated by 270 degree and appears to be horizontally oriented. Is therer any specific reason for this?

Thank you for your time and patience to answer my questions. :)

stephanecharette commented 3 years ago

My guess is the image has some embedded meta information about the orientation, and OpenCV which is used to load images either does or does not follow that meta information. My guess is if you used something like DarkMark to look at your annotations, you'd find that (some of) the annotations are not correct, which is causing the problems you're seeing.

AkkiSony commented 3 years ago

@stephanecharette Thanks. I'll look into the annotated images once. :)