problem on landmarks shift

[PandaGuy0623]

Hi, I have this problem when I run your demo. The following picture shows that when I run the demo, the landmarks have a great shift from where they ought to be. However, if I stream the image to the host and use mediapipe solution directly, the result is good.

I wonder if this problem is related to the blob file generated from tflite? How can I solve this problem? Thanks.

[geaxgx]

Yes, you are right, we have lost in accuracy when converting from tflite to blob. It happens for some models (similar problem with movenet). Here, one possible explanation is that we have less precision with float16 than in the original model (https://github.com/PINTO0309/tflite2tensorflow/issues/9#issuecomment-842460014). Another possible cause: https://github.com/openvinotoolkit/openvino/issues/7674 So there is not much we can do about it. IIRC, the landmarks are less shifted with the 0.8.4 version of the model. You can try by using the older branch mediapipe084 (git checkout mediapipe084). Otherwise, google should release (soon ?) a new version of their blazepose model, if I refer to this tweet: https://twitter.com/yeemachine/status/1539639359657447424

[PandaGuy0623]

Thanks for your reply!!

Yes, you are right, we have lost in accuracy when converting from tflite to blob. It happens for some models (similar problem with movenet). Here, one possible explanation is that we have less precision with float16 than in the original model (PINTO0309/tflite2tensorflow#9 (comment)). Another possible cause: openvinotoolkit/openvino#7674 So there is not much we can do about it. IIRC, the landmarks are less shifted with the 0.8.4 version of the model. You can try by using the older branch mediapipe084 (git checkout mediapipe084). Otherwise, google should release (soon ?) a new version of their blazepose model, if I refer to this tweet: https://twitter.com/yeemachine/status/1539639359657447424

[tristanle22]

@PandaGuy0623 Have you been able to get around this issue?