imx93 dms face_landmark model, the output of the model is not consistent with the output processing

cancel
Showing results for 
Show  only  | Search instead for 
Did you mean: 

imx93 dms face_landmark model, the output of the model is not consistent with the output processing

Jump to solution
928 Views
DOMINICK
Contributor II

The face_landmark model downloaded by download_models.py in /usr/bin/eiq-examples-git of the im93 sdk version 6.1.55 is inconsistent with the output processing of the routine. If possible, please provide a model that fits this routine

DOMINICK_0-1704696309234.png

 

0 Kudos
1 Solution
840 Views
Zhiming_Liu
NXP TechSupport
NXP TechSupport

你用这个里面的模型,不要用新的。

View solution in original post

6 Replies
889 Views
Zhiming_Liu
NXP TechSupport
NXP TechSupport

The url is not maintained by NXP, you can refer this github link:

https://github.com/PINTO0309/PINTO_model_zoo/blob/main/043_face_landmark/download.sh

0 Kudos
867 Views
DOMINICK
Contributor II

你好,我明白不是nxp维护的这个模型,只是现在imx93里面该目录下的推理例程和下载下来的这个模型不匹配了(可能这个网站做了模型迭代导致的),麻烦反馈提供新的模型或者新的推理例程或者修改方法,现在是模型的输出和例程输出完全不匹配直接会导致推理错误。

0 Kudos
859 Views
Zhiming_Liu
NXP TechSupport
NXP TechSupport

你用的是什么命令测的?CPU推理还是NPU?

0 Kudos
854 Views
DOMINICK
Contributor II

1、使用的是cpu推理,执行命令python3 main.py -i /dev/video1      (我的usb摄像头是video1)

2、使用下载的landmark的模型根据下载脚本提供解压后的resources\10_new_detector_landmark\resources\saved_model里面的model_float32.tflite,我先把模型名字改成了和例程一致的,然后发现输出实际不一致,通过netron查看,发现和例程要求的名字不一样,例程要求的是2输出(可以看face_landmark.py里面),但实际模型是3输出

 

麻烦实际测一下就知道问题所在了

DOMINICK_0-1705040666949.png

DOMINICK_1-1705041563864.png

DOMINICK_2-1705041682383.png

 

 

 

0 Kudos
841 Views
Zhiming_Liu
NXP TechSupport
NXP TechSupport

你用这个里面的模型,不要用新的。

836 Views
DOMINICK
Contributor II

可以用了,谢谢

0 Kudos