-1

I am making a project that utilizes MLKit. The classification model will be a TensorFlow Lite model. I noticed that the detected objects always return rectangular bounding boxes. I would like them to return polygonal bounds that are shaped like the object it is detecting, or if possible, a sort of "3D" bound.

I am aware of certain annotation tools, along with things like Mask RCNN, but I am not sure how to integrate them into a TensorFlow Lite model (or if I am supposed to implement it in the model rather than the base code) or if I can even do it at all, hence why I am asking...

Is it possible to make the detected objects return bounding polygons, or even 3D polygons/image segmentations, instead of bounding boxes, using MLKit + TensorFlow Lite?

Erick Adam
  • 399
  • 6
  • 13

1 Answers1

0

ML Kit's Object Detection & Tracking API is limited to 2D bounding boxes for detected objects. I am personally not aware of generalized object detection solutions that return polygons as boundaries. Image segmentation (e.g. Tensorflow Lite Image Segmentation) seems like an easier approach, although you may need to (re-)train your own model depending on the application.

Chrisito
  • 494
  • 2
  • 4