Closed russelldj closed 1 month ago
I also implemented NMS, which was fairly straightforward. One major limitation is this library only supports rectangular bounding boxes. Currently, this is defined as the minimum bounding rectangle for each polygon (.bounds
). In the future we might want to support real polygon NMS. Also, there's an option to include a multi-class variation in LSNMS
, the library we use. I didn't implement it, but we could consider it in the future.
Most of this PR is making changes/extensions to the
RegionDetections
andRegionDetectionsSet
classes to support streamlined NMS.merge
method toRegionDetectionSet
to handle the logic of merging multiple regions together or returning a list of per-region detections.get_detections
toget_data_frame
RegionDetections
with a subset of detections.RegionDetections
more flexible so inputs can be provided in geospatial coordinates but also have a transform to pixels associated.Also I currently have un-used functionality to get the bounds of the predicted region in different CRS for both the
RegionDetections
andRegionDetectionSet
.