Open aching1 opened 2 years ago
Initial idea for implementation: add a config for "pairable" features. Add all raw boxes as an optional input to process_raw_box. During process_raw_box, if a label is pairable, calculate distance metric to nearest raw box above threshold, and if less than the critical distance, decrease the min_prob for promoting the raw box to a censorable box.
Better implementation would be to only deal in pairs, but that would require some much more complicated work and almost surely not worth it.
It's often possible to create a scenario whereby one of a paired item (breasts, feet etc) is detected. In such a case, we have an improved prior that a second is likely to be present, but wasn't meeting the confidence threshold, so we'd like to try searching to find it with lower threshold. Some possible approaches: