SOS! TensorRT engine different infer output in yolov4 on DeepStream5.0

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson / GPU) GTX1660s
• DeepStream Version deepstream5.0
• TensorRT Version 7.1
• NVIDIA GPU Driver Version (valid for GPU only) 450.51

Use GitHub - Tianxiaomo/pytorch-YOLOv4: PyTorch ,ONNX and TensorRT implementation of YOLOv4 to convert .weights to onnx and engine files for testing,
where:

  1. Via GitHub - Tianxiaomo/pytorch-YOLOv4: PyTorch ,ONNX and TensorRT implementation of YOLOv4
    The script in onnx and engine are tested, and the results can be consistent.
  2. Integrate the engine generated above into deepstream5.0, most of the results can be consistent, but some more target boxes will be detected on some images.
    That is, some false positives will be generated when passing deepstream detection.

Excuse me, is there any solution?

Thank you!

3

I made a workable YoloV4 version (here ) based on the GitHub - NVIDIA-AI-IOT/yolov4_deepstream project .

Could you reproduce the issue with aove YoloV4 DS and share the screenshot?

4

We use custom data, how do you test it?

5

This issue is observed on your side, what do you ask us to test?

In my above reply, I mean you reproduce the issue with aove YoloV4 DS and share the screenshot of the issue - some false positives will be generated when passing deepstream detection?

6

I can provide models and .engine files. And test images, can you confirm our problem. Because we have been troubled by this problem for many days.

7

Or whether other customers also encountered the problem, we always inconsistent with our model test results in the actual product deployment environment, which caused us a lot of trouble.

8

I mean I have a workable YoloV4 version (here ), you can use your custom data to test. We have verified the mAP of this model.
If you can’t reproduce the issue with this DS app & model, it indicates the issue comes from your model, you need to retrain your model.

9

I know, maybe I am not clear about it. I have downloaded it many times, but it will be aborted every time. Are there other ways to download the above files?

10

The .engine file we generated, through GitHub - Tianxiaomo/pytorch-YOLOv4: PyTorch ,ONNX and TensorRT implementation of YOLOv4

in:
python3 demo_trt.py yolov4.engine 0.jpg 416 416

The same result as the original model can be obtained.
But it can’t with deepstream5.0.

11

Ok, so it may be related to the pre-or poset-processing of the non-DS solution.
If it’s, it’s not a DS related issue, we can’t help it, sorry!