Is it possible to train ACF detector with polygon labeled data?

Hi, maybe this is so simple question that i can not find any information regarding this issue.

I use image labler app to define ground truth training data for ACF detector. During the process i noticed that if my object of intrest is turned by some angle (for example 45deg.) by using rectangle i am capturing a lot of misleading information for the detector. Which results that detection confidence score is very low, and some times it misses the defect in question all together.

I was trying to use polygon to define these objects for training, but then i try to train ACF detector i have error that ground truth data is empty. So, i am wondering does ACF supports polygon labels ? And if so, how to use them ? Or maybe i need to use another type of detector ?

If not maybe you can share your experiance for better approuch to define such objects for detector training ?

Answer 1

[Vandit]

Hi,

Yes, ACF supports polygon labels. To use them, you need to convert the polygon labels to a format that ACF can understand. Below are the steps on how to use polygon labels to train an ACF detector:

1. Use an image labeling app to label your images with polygons.
2. Convert the polygon labels to masks using the 'poly2mask' function.
3. Convert the masks to rectangles using the 'poly2rect' function.
4. Train the ACF detector on the images and masks using 'trainAcfDetector' function.
5. Once the ACF detector is trained, use 'detectAcfDetector' function in MATLAB to detect objects in new images.

If the ACF detector does not produce satisfactory results with polygon labels, you can consider using other object detection algorithms in MATLAB, such as the 'YOLO v2' or 'Faster R-CNN' detector, which may perform better with polygon-labeled data.

To know more about the above mentioned MATLAB functions, refer to the link below:

1. Convert region of interest (ROI) polygon to region mask - MATLAB poly2mask (mathworks.com)
2. Train ACF object detector - MATLAB trainACFObjectDetector (mathworks.com)
3. Detect objects using ACF object detector - MATLAB detect (mathworks.com)

Hope this helps.

Thankyou