'Export TFlite ssd-mobilenet without NMS and with bounding boxes decoding logic using Tensorflow Object Detection API
I converted an ssd_mobilenet_v1 model in TFlite format using the Tensorflow Object Detection API, with the export_tflite_ssd_graph.py. Since I don't want the post-processing (NMS) in my final graph, I set the parameter --add_postprocessing_op to false. The exported model has two outputs raw_outputs/box_encodings and raw_outputs/class_predictions. In particular the raw_outputs/box_encodings contain raw bouding boxes, which need to be decoded using the anchor boxes, as explained here.
Is there a way to export this model, maintaining the bounding box decoding functionality within the graph?
Sources
This article follows the attribution requirements of Stack Overflow and is licensed under CC BY-SA 3.0.
Source: Stack Overflow
| Solution | Source |
|---|