Using a trained Faster RCNN Object Detector, expected LABEL to be nonempty

I’m using https://www.mathworks.com/help/vision/examples/object-detection-using-faster-r-cnn-deep-learning.html example to train it in my dataset.

the dataset contains 400 images and after it finish’s training "expected LABEL to be nonempty" error occur

What can I do?

Answer 1

[Sourav Bairagya]

It seems that the labelling is not done properly.

It may be possible that some images are not labelled. So, it would be good to check the labelling of the input dataset once.

There is another possibility that if the bounding boxes present in input image are already small, then after resizing them, those small bounding boxes will become more smaller. Hence, while processing through first few layers of resNet50, all the useful labelling informations will be lost due to further downsampling of already small bounding boxes. To avoid such incident, you can crop the desired portions from the original image and adjust the bounding boxes accordingly so that after resize they don't become too small.

[jaida alsuhaibani]

Thank you for your answer I guess the problem is that the bounding boxes sometimes are really small Do you have any suggestions of how we can fix it without manually checking the images?