autoML vision missing training bounding boxes

Question

I am trying to create a new object detection model in autoML vision. I followed this & this guides about how to prepare and format my training data. For some unknown reason, there are a lot of missing bounding boxes when importing the data. For example, an image with 84 bounding boxes only loads 12 in autoML.

I have checked for the minimum bounding box size, maximum number of bounding boxes per image and maximum image size.

Is anybody else experiencing the same issues?

Answer 1

There are two additional requirements that you didn't mention in the description, and I'm adding them here, just in case.

• Bounding box edge length: At least 0.01 * length of a side of an image. For example, a 1000 * 900 pixel image would require bounding boxes of at least 10 * 9 pixels

• All the bounding boxes should be inside images

Also, there seems to be some limits regarding bounding boxes and labels in the UI; however, the limit is supposed to be 50, and 12 is really far from that.

If you are sure that your bounding boxes, as well as the CSV, satisfies all the requirements, I suggest to open a new Issue Tracker to the Cloud Vision team, so they can take a deeper look at your issue.