Hey there,
I'm trying to deploy a face detection tflite model on AM62A. The two outputs of my model are the score of face existence in the image and 6 approximate facial keypoint coordinates plus facial bounding box coordinates.
I set 'object_detection:meta_arch_type' to 1 and 'model_type' to 'OD'. Then I try to write the prototxt file for 'object_detection:meta_layers_names_list'. If I understand correctly, I can put the post-processing of my SSD model on C7X through this method? I check the official example in https://github.com/TexasInstruments/edgeai-modelzoo/blob/master/models/vision/detection/coco/tf1-models/ssdlite_mobiledet_dsp_320x320_coco_20200519.prototxt.link. I am not sure why there is the 'faster_rcnn_box_coder' parameter. For my SSD model, it directly output the offset of bounding boxs relative to default anchors. Do I really need to set this parameter in my prototxt file? I upload my model and my script in this link: https://www.dropbox.com/scl/fi/lvv2g9917xh7atunzkpht/face-detection.zip?rlkey=s5l9sh6qohv9jjf1ogejiowap&dl=0.
Secondly, should my model be the exactly original SSD model for the post-processing optimization? Could there be any modification of SSD model?
Thirdly, I notice there is one custom operator TFLite_Detection_PostProcess in the example model. In order to utilize C7X for post-processing, should I also add this operator inside my model? If my model just ends up with the reshape and concatenation operators, and with the prototxt file, can I also put my post-processing inside C7x? IF NOT, how can I add this custom operator in my code? I don't know where this operator comes from.
Thanks for you help.
BR,
Yunfeng
