Discrepancies in Object Detection Results between Ultralytics and DeepStream at Yolov8

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

• Hardware Platform (Jetson / GPU): dPU A40.
• DeepStream Version: 6.4-triton-multiarch.
• TensorRT Version: 8.6.2.3.
• NVIDIA GPU Driver Version (valid for GPU only): 525.147.05.
• Issue Type( questions, new requirements, bugs): questions.
• How to reproduce the issue ? (This is for bugs. Including which sample app is using, the configuration files content, the command line used and other details for reproducing)

I am using DeepStream with Python, specifically the repository GitHub - marcoslucianops/DeepStream-Yolo: NVIDIA DeepStream SDK 6.3 / 6.2 / 6.1.1 / 6.1 / 6.0.1 / 6.0 / 5.1 implementation for YOLO models to convert a YOLOv8x.pt model to ONNX for use in a single-class person detection pgie model. I’ve noticed significant differences in the results when comparing the model’s performance using Ultralytics for inference versus using DeepStream. In Ultralytics, I set the iou to 0.7 and the threshold to 0.5, and similarly in DeepStream, I set the nms-iou-threshold to 0.7 and the pre-cluster-threshold to 0.5. Am I configuring something incorrectly? Can you identify why there might be such a significant discrepancy in the results between the two methods?
Below are the results of using ultralytics for inference:

204
2041280×720 330 KB

And here is the result from deepstream:
Screenshot 2024-02-06 at 21.45.53
Screenshot 2024-02-06 at 21.45.531920×1076 198 KB

Here is the ultralytics code I am using for inference, config pgie deepstream, model pt and model onnx:
yolov8 - Google Drive

3

Could you attach the original image to us? And how to run that with Ultralytics?

6

I uploaded everything to this link: yolov8 - Google Drive. 204_origin is the original image, 204_ultralytics is the inference image using ultralytics, 204_deepstream is the result I get when using deepstream and built_in_infer.py is the script for inference using ultralytics. Please help me, thank you very much.

7

DId you implement the NvDsInferParseYolo yourself? We have summarized some experience in accuracy. Could you refer to the Debug Tips for DeepStream Accuracy Issue and check if that helps?

8

@taolanhat were you able to find a solution for the issue? I’m having an identical issue using the following nvinfer config.

9

Found the following git issue from DeepStream-Yolo repository owner.

Apparently, TensorRT drops the accuracies when converting model, particularly, during layer optimizations.

The issue is still open and I couldn’t find any resolution in the shape of a fix, alternative or workaround.