Faciallandmark model only works when there is only one face in the image

Accelerated Computing Intelligent Video Analytics DeepStream SDK
, ,
1

Hello,

With my team we are working on a Computer Vision project and for our solution, we are using NVidia tools and more precisely for this part the FacialLandmarks and the Gaze models. But we are facing a problem when executing the app : it works well when there is only one person on the camera but when there is more persons, the landmarks and gaze are correctly detected only for the first person seen on the video while they appear totally random for the other persons.

Here is our configuration :

Hardware platform : Jetson AGX ORIN DevKit

Deepstream 7.0

Jeckpack 6

TensorRT 8.6.2.3

Cuda 12.2

NVIDIA GPU Driver version 540.3.0

Have you seen this behavior on your side ?

3

Are you referring to our deepstream-gaze-app demo? And could you attach your video source to us?

4

Yes it is indeed the deepstream-gaze-app demo. Video source is the camera Panasonic AW-UE4KG

5

Plus we are using /dev/video0 to get the video stream

6

I have tried that on my side. It can detect more than 1 face. Could you dump the video source from your v4l2 camera and attach to us?

7

The face detect is also working on our side (we can have more than one red boxes), but the landmarks inside these boxes seems to be random. Are you able to get the landmarks even on multiple faces ?

8

We are analyzing this issue and will inform you promptly when we have a conclusion. Thanks

1 Like
10

Do you have reproduced the issue?

11

Yes. We’re analyzing it internally.
@sgalibert @msandjivy Could you try to use the fp16 mode and see if there is still a problem?

  1. change the network-mode to 2.
  2. change the model-engine-file to ../../../models/faciallandmark/faciallandmark.etlt_b32_gpu0_fp16.engine
2 Likes
13

Yes it fixes it. thanks

14

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.