Facial Landmark Estimator (FPENet); how to use on 1920x1200 resolutuion

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

• Hardware : Jetson Orin Nano
• Network Type (Tao Facelandmark FpeNET)
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here)

My Input from the camera is 1920x1200. I want to use on this use facelandmarks FPEnet and crop eyes seperately without loosing resolution but the max input tensor sgie is = 80x80 and jetson max scaling = x16 therefore x1/16; 80*16=1280 . Is there any way that I can use 1920 resolution?? Appreciate your help, best regards

3

What do you mean by “jetson max scaling”?

For running fpenet, you can refer to the pipeline in
https://github.com/NVIDIA-AI-IOT/deepstream_tao_apps/tree/master/apps/tao_others/deepstream-faciallandmark-app and config in https://github.com/NVIDIA-AI-IOT/deepstream_tao_apps/tree/master/configs/nvinfer/facial_tao. The 1920x1200 should be working.

4

Thanks for your reply. If the PGIE forwards a picture bigger than 1280 size, we receive this error:

/dvs/git/dirty/git-master_linux/nvutils/nvbufsurftransform/nvbufsurftransform.cpp:4575: => VIC Configuration failed image scale factor exceeds 16, use GPU for Transformation

We needed 24 on jetson orin. any suggestions?

5

Could you try to refer to the solution mentioned in DeepStream SDK FAQ - #29 by fanzh ?

6

I need to scale more than 16. I will try scaling-compute-hw=1 .
Would rescaling the pgie obj_meta rect_params distort the results in shie?

Any other suggestions?

7

According to DeepStream SDK FAQ - #29 by fanzh, can you add below into https://github.com/NVIDIA-AI-IOT/deepstream_tao_apps/blob/master/configs/nvinfer/facial_tao/config_infer_primary_facenet.txt and retry?

input-object-min-height=26
input-object-min-width=46
1 Like
8

1920x1200 input -->PGIE(facedetect)–>SGIE(facelandmark)
PGIE(facedetect):
infer-dims=3;736;416
input-object-min-height=26
input-object-min-width=46

This does not help as the scaling is not there; I do not receive small faces so that PGIE has to find more than 16 times smaller faces. I had scaling problem at SGIE, where the frame was more than16x bigger than infer-dim of SGIE. I can limit in opposite input-object-max-height or not use VIC and use GPU. I hope this does not bring to much work for GPU, for that I have to profile. At the Facelandmark example, it has virtually no effect, both have about 57 FPS. I appreciate your helps, thanks, it is solved now

9

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