Please provide the following information when requesting support.
• Hardware v100
• Network Type fpnet
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here) 3.0
• Training spec file(If have, please share here)
This is not an issue, it is a requirement,
Like the detection and classification models in tao tookit, We can use transfor
learn method to adapt to other new object class’s detect and classification,
But target to Landmarks(keypoints) Estimation, currently fpnet is very limitations, Ii can only transfor learned
for only face and only support 68, 80 104 keypoints.
I think in computer vision applications, not only the face need the keypoint estimation, many other object may also
need the keypoint estimations according to the use case, and the need estimationed keypoints may be not 68/80/104; according to the use case, for example the object only need to estimate 6 or 11 keypints.
So I think advise you to update this fpenet to be more gerneral, to make it can transfor learned for any objects(not
only face) and any number of keypoints(not only 68 80 and 104), and the model can be renamed as gerneral_object_keypints_estimation_model.
thank you!