TensorRT Initialization Failure

[sayyid-abolfazl]

Hello guys, thank you all for everything.

I am trying to use the s_face face_recognition model in tensorRT mode :

https://github.com/opencv/opencv_zoo/tree/main/models/face_recognition_sface

To start, I checked the model with the netron tool to get the model inputs and outputs :

Now, let's compare two photos for a test :

This is my program at this address
https://github.com/sayyid-abolfazl/sface_trt

and its output is not correct

The problem is that the feature vectors obtained in the onnx model are completely different from the Nvidia TRT version, and I don't know where the problem lies.

https://github.com/opencv/opencv_zoo/blob/main/models/face_recognition_sface/demo.cpp

It should be noted that if you use the align and crop method, the results are slightly different from the case where you just send the cropped face image without alignment, but in the TRT the results are completely different.

I tried several different methods, such as changing the color channel or the order of the channels and how the network blob was created, but I didn't get the right output.

I am trying to use the model's own alignment and cropping mode and send the image with 15 facial coordinates to it, but currently the output in normal mode is completely wrong.

[fengyuentau]

Hello, TensorRT is not supported by OpenCV. For helpful and faster responses, you should go to https://www.nvidia.com/en-us/forums/.

[Mo-Sam-Mo]

Hi,
I’d like to help with this issue. Could you provide more context or guidance on where to start?

[sayyid-abolfazl]

Hi, I’d like to help with this issue. Could you provide more context or guidance on where to start?

Hello, Thank you for your support.

To start, please run my program, which is based on the Tensor RT library
https://github.com/sayyid-abolfazl/sface_trt

to use the onnx model in Tensor RT, you need to check the input and output of the network and then implement it in Tensor RT.

The output of the TRT model is completely different and I don't know what the problem is.

I used fp16 - fp32 - best modes to get TRT output but the results are still wrong.