PROCESSOR-SDK-AM69A: Human Pose Export Issue

Hi Bill,

Please reference the following edgeai-tidl-tools document that explains object detection (OD) heads and meta architectures with prototxt examples: https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_fsg_od_meta_arch.md

Let me know if you have any further questions about this.

Thank you,

Fabiana

Hi Fabiana. Thank you for your response. 
I have checked the reference you shared for object detection. However, I am still a little bit confused, when I checked original model from TI for Keypoint detection, I think prototxt for Keypoint Detection is somehow different, there are "num_keypoint" and "keypoint_confidence" value, those values are not on object detection prototxt, Could you explain about this ? and if I have YoloV5 Pose model, and create object detection prototxt without "num_keypoint" and "keypoint_confidence", will my model and artifacts still able to perform keypoint detection task ?   
Thank you before and looking forward for your response. 

Hi Bill,

Do you have a pt file as part of your ONNX export? There is a command that generates the required prototxt file using the pt file. This information is outlined here: https://github.com/TexasInstruments/edgeai-yolov5?tab=readme-ov-file#onnx-export-including-detection

I have looped in one of our Keypoint Detection experts to answer your question. In the meantime, an explanation of YOLOX Keypoint Detection is available here: https://github.com/TexasInstruments/edgeai-yolox/blob/main/README_keypoint_detection.md

Thank you,

Fabiana

Here is a model and prototxt that was trained in edgeai-yolox

https://github.com/TexasInstruments/edgeai-tensorlab/tree/main/edgeai-modelzoo/models/vision/keypoint/coco/edgeai-yolox

This will help you understand the structure of the prototxt.

edgeai-yolox is here:

github.com/.../edgeai-yolox

Hi Fabiana, Thank you for your response and assist, I really appreciate it. 
However using this https://github.com/TexasInstruments/edgeai-yolov5?tab=readme-ov-file#onnx-export-including-detection repository, I can generate prototxt for Yolov5-pose detection but without keypoint information ("num_keypoint") on prototxt file. Which I believe this information might be very important to human pose detection ?   Looking forward for your response and thank you before 

Hi Manu, Thank you for your response and assist, However I still have another question. 
How If I want to use yolov5s-pose from this repository (https://github.com/TexasInstruments/edgeai-yolov5?tab=readme-ov-file#onnx-export-including-detection), How can I generate prototxt file ? 
the export program seems to not include the function to generate prototxt which I think it is needed, do I have to do it manually based on this (https://github.com/TexasInstruments/edgeai-tensorlab/tree/main/edgeai-yolox) ? 
Looking forward for your response, thank you before. 

Hi Manu. Thank you for your response. 
Thank you once again for good explanation I can totally understand now and I know how to solve my current issue. 
If I want to manually create protoxt file, can the documents below be used as references for manually creating prototxt file?
https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_fsg_od_meta_arch.md 
https://github.com/TexasInstruments/edgeai-tensorlab/tree/main/docs/tech_reports/20240901-speed-accuracy-analysis-of-object-detection-models#performance-on-tidl

Yes - this link has several example prototxt files and the explanation there will help you understand how the prototxt file is to be written.