Landmarks spread over face in TAO facial landmark sample app

[aniketrohara]

• Hardware: Jetson AGX Orin
• Deesptream7.0
• Jetpack 6.0 (6.0+b106 and 6.0+b87 both are installed) L4T 36.3.0
• TensorRT 8.6.2
• NVRM version: NVIDIA UNIX Open Kernel Module for aarch64 540.3.0
• Issue : output landmarks are spread all over the face and are improper as shown below, (i am using the sample application given for facial landmarks on my input video with multiple faces in a single frame)
• models: default models provided (i am NOT using any custom model)
  
  

[fanzh]

did you modify the code? seems it is model accuracy issue. you can retrain the model. please find " How to use this model" in this link.

[aniketrohara]

But i have not modified the code so still is there any need to retrain the model?..if yes, so does this mean that the provided model is not that much accurate?

[fanzh]

we tested the app with some faces. the facial landmarks are correct. AKY, we can’t share faces due to the legal reasons. maybe you can try a big frontal face or you can retain the model to improve the accuracy.

[aniketrohara]

okay so do you think that the issue could be because of my input video resolution?

[fanzh]

yes. resolution, light, angle, etc. will all affect the model accuracy.