Mask out bounding boxes from other detections in frame

[ejm714]

For distance estimation, if there are multiple detections in the frame, the distance estimate is repeated (but the same for both). This is because the input used for each is the full frame.

To generate predictions that are distinct per detection, we could consider masking out the other detections (e.g. setting to zero). This would benefit from some evaluation in a notebook first to determine if masking throws off the model predictions altogether or is helpful.