Support for per-class NMS score thresholds.

Files

https://github.com/tensorflow/models/blob/master/research/object_detection/protos/post_processing.proto
https://github.com/tensorflow/models/blob/master/research/object_detection/export_tflite_ssd_graph.py
https://github.com/tensorflow/models/blob/master/research/object_detection/export_tflite_graph_tf2.py

Feature Request

Ability to specify a separate NMS score threshold for each class when exporting a model to TFlite.

Context

In multi-class detection tasks, the optimal score threshold may vary significantly between classes (i.e. one may require a threshold of 0.4 for stop signs and 0.6 for vehicles). Currently, post-processing (batch NMS) only allows the user to specify one universal threshold for all detections. This leaves the user with only 2 options: either compromise on overall accuracy, or to write & optimize code to handle per-class post-processsing.

Are you willing to contribute?

Yes (though may not have the expertise)