Pixel labeling for Training RCNN detector

Hi Guys

Just asking Is it possible to make Pixel labeling by groundTruth for RCNN Detector

Answer 1

[Raunak Gupta]

Hi,

I am assuming pixel labelling here means to doing Semantic Segmentation of the image you may be working on. Since rcnnObjectDetector is used specifically for Object Detection in a image and Object classification it cannot be used for doing semantic segmentation. You may get started with Semantic Segmentation using Deep Learning for pixel labelling.

For learning more about semantic segmentation you may refer to the following:

https://www.mathworks.com/solutions/image-video-processing/semantic-segmentation.html

https://www.mathworks.com/help/vision/ug/semantic-segmentation-examples.html

You may refer to unetLayers and segnetLayers as these are generally used as a headstart in Semantic Segmentation.

[Abdussalam Elhanashi]

Hi Raunak

Thanks for the answer

just asking, even i used the line to draw irrigular shape which this feature is present in the app however it does not work when i load the gTruth to workplace i checked the variables it shows as cell and not like coordinates numbers as done with rectangular do you have explaination on that?

Regards

[Raunak Gupta]

Hi,

For Object Detection problems, bounding boxes are generated as an end result from the network. So, drawing irregular shapes may not help as the bounding boxes are rectangular in nature specifying only the corner points. I assume this is the problem you may be facing as irrelgular shape will generate several set of coordinates which generally don't represent rectangle. It is recommended to use rectangular bounding boxes for training in such cases.