Facial Landmarks NVIDIA

[melissa1]

Hello, I am working on Jetson AGX Xavier, using deepstream-5.0
I am probing the data from the secondary of the facial landmarks.
However, I am getting fuzzy output. The output is really small. Here’s an example of the output:
[[ 0.3544922 1.0419922 ]
[-0.86279297 0.41308594]
[ 0.70654297 0.6381836 ]
[ 1.5234375 0.5991211 ]
[ 0.29882812 -1.5205078 ]
[ 1.2011719 -0.67871094]
[-0.38964844 2.3886719 ]
[ 1.2324219 -0.32421875]
[ 1.4072266 -0.8071289 ]
[-1.4365234 -2.0683594 ]
[-2.3027344 -0.7402344 ]
[ 0.5625 -0.5751953 ]
[-0.3486328 -0.6738281 ]
[ 1.0292969 -1.015625 ]
[ 0.36108398 0.18859863]
[-0.49365234 -1.1875 ]
[ 1.8046875 -0.42871094]
[ 2.0566406 0.5126953 ]
[-1.9404297 0.29516602]
[-1.171875 0.6591797 ]
[-0.73095703 -0.44091797]
[-0.26708984 -0.7993164 ]
[ 0.6123047 -2.2578125 ]
[-0.9169922 0.8725586 ]
[ 1.0371094 0.01264954]
[ 0.6748047 0.71240234]
[ 0.03775024 -0.7758789 ]
[ 1.3359375 1.0849609 ]
[ 1.4960938 0.43115234]
[-0.7919922 -1.2880859 ]
[-1.015625 1.2324219 ]
[ 2.0234375 1.9335938 ]
[-0.15686035 0.9121094 ]
[-1.9501953 1.0009766 ]
[ 0.19909668 0.76416016]
[ 0.9238281 -0.86083984]
[ 1.0205078 -0.7988281 ]
[ 0.5961914 -0.43945312]
[-0.17602539 -0.87109375]
[-0.03256226 -0.9404297 ]
[ 0.54345703 2.1933594 ]
[-1.1660156 -1.0751953 ]
[ 1.4921875 0.35913086]
[-0.09942627 -0.5527344 ]
[ 0.18432617 -0.453125 ]
[-0.21533203 2.5839844 ]
[ 1.9882812 0.22045898]
[-1.1152344 2.2246094 ]
[-0.6870117 1.1054688 ]
[-2.8828125 0.9609375 ]
[-1.5693359 -0.18920898]
[ 0.37109375 1.4570312 ]
[ 2.6308594 -1.1152344 ]
[ 0.2705078 0.03457642]
[ 0.07250977 0.10241699]
[-0.7832031 1.5634766 ]
[-0.28271484 1.0146484 ]
[-2.1464844 -0.82470703]
[ 0.58740234 0.3322754 ]
[ 1.9140625 -0.37646484]
[-0.5810547 0.5288086 ]
[ 1.5693359 -1.4111328 ]
[ 0.7001953 1.5712891 ]
[-0.5151367 -0.16992188]
[ 0.22631836 1.0009766 ]
[-0.93066406 0.5048828 ]
[ 0.7626953 0.55371094]
[ 1.7714844 0.3876953 ]
[ 0.27929688 -1.6728516 ]
[ 1.2871094 -0.7392578 ]
[-0.38378906 2.3378906 ]
[ 1.1855469 -0.6435547 ]
[ 1.4267578 -0.7705078 ]
[-1.5087891 -2.0585938 ]
[-2.2070312 -0.8276367 ]
[ 0.59277344 -0.8417969 ]
[-0.42797852 -0.5151367 ]
[ 1.0087891 -1.3388672 ]
[ 0.17407227 0.1583252 ]
[-0.41210938 -1.015625 ]]

Is there a way to solve this issue?

[bcao]

Please check the preprocess of your model and the nvinfer configs for the model.

[melissa1]

I am generating the output, the 80 facial landmarks in the secondary.
However, I was checking the following link:

Understanding settings for secondary classifier

Thanks for your answers, I was thinking of your comment “Deepstream is only a SDK, you need to understand the interfaces and usages with document and source code.” I almost agree with you, but In my view i was hoping that the documentation should be enough to be able to use existing components. For example the “gst-nvmsgconv” why should I read the source-code to change between different message-formats? Why is there no link to the open-source page of that module? Another example, the “infer-…

There’s a parameter called num_keypoints to be specified. What are the other specs to be specified in the config file?
I am stuck with it because there isn’t any documentation about the secondary in the drowsiness model.

[bcao]

@melissa1

There’s a parameter called num_keypoints to be specified. What are the other specs to be specified in the config file?

I think num_keypoints is the item from spec file which is for TLT model training, you can create a new topic in TLT forum for more specific info.
If you want to deploy the model to deepstream, currently we only support standard dectction, that means we only can generate bbox in the post parser. If you think that cannot meet your requirement, then you can try to attach the raw tensor to the meta data, you can refer deepstream_infer_tensor_meta_test.cpp