README.en.md

简体中文 | English

Instance Segmentation Inference Example

This example uses the YOLO11n-seg model to demonstrate how to perform Instance Segmentation inference using the Command Line Interface (CLI), Python, and C++.

yolo11n-seg.pt，【TestImages】COCO-part.zip

Please download the required yolo11n-seg.pt model file and test images through the provided link, and save the model file to the models folder, and place the extracted test images into the images folder after unzipping.

Model Export

Important

If you only want to export the ONNX model (with TensorRT plugins) that can be used for inference in this project through the tensorrt_yolo provided Command Line Interface (CLI) tool trtyolo, you can install it via PyPI by simply executing the following command:

pip install -U tensorrt_yolo

If you want to experience the same inference speed as C++, please refer to Install-tensorrt_yolo to build the latest version of tensorrt_yolo yourself.

Use the following command to export the ONNX format with the EfficientRotatedNMS plugin. For detailed trtyolo CLI export methods, please read Model Export:

trtyolo export -w models/yolo11n-seg.pt -v yolo11 -o models -s

After running the above command, a yolo11n-seg.onnx file with a batch_size of 1 will be generated in the models folder. Next, use the trtexec tool to convert the ONNX file to a TensorRT engine (fp16):

trtexec --onnx=models/yolo11n-seg.onnx --saveEngine=models/yolo11n-seg.engine --fp16 --staticPlugins=/path/to/your/TensorRT-YOLO/lib/plugin/libcustom_plugins.so --setPluginsToSerialize=/path/to/your/TensorRT-YOLO/lib/plugin/libcustom_plugins.so

Model Inference

Important

The tensorrt_yolo installed via PyPI only provides the ONNX model (with TensorRT plugins) for inference in this project and does not provide inference capabilities. If you want to experience the same inference speed as C++, please refer to Install-tensorrt_yolo to build the latest version of tensorrt_yolo yourself.

Inference Using CLI

Note

The --cudaGraph command added from version 4.0 can further accelerate the inference process, but this feature only supports static models.

From version 4.3 and later, support for Instance Segmentation inference is added. The command -m 2, --mode 2 is used to select the Instance Segmentation.

1. Use the trtyolo command-line tool from the tensorrt_yolo library for inference. Run the following command to view help information:

   trtyolo infer --help

2. Run the following command for inference:

   trtyolo infer -e models/yolo11n-seg.engine -m 1 -i images -o output -l labels.txt --cudaGraph

   The inference results will be saved in the output folder, and a visualization result will be generated.

Inference Using Python

1. Use the tensorrt_yolo library to run the example script segment.py for inference.

2. Run the following command for inference:

   python segment.py -e models/yolo11n-seg.engine -i images -o output -l labels.txt --cudaGraph

Inference Using C++

1. Ensure that the project has been compiled according to the TensorRT-YOLO Compilation.

2. Compile segment.cpp into an executable:

   # Compile using xmake
   xmake f -P . --tensorrt="/path/to/your/TensorRT" --deploy="/path/to/your/TensorRT-YOLO"
   xmake -P . -r
   
   # Compile using cmake
   mkdir -p build && cd build
   cmake -DTENSORRT_PATH="/path/to/your/TensorRT" -DDEPLOY_PATH="/path/to/your/TensorRT-YOLO" .. 
   cmake --build . -j8 --config Release

   After compilation, the executable file will be generated in the bin folder of the project root directory.

3. Run the following command for inference:

   cd bin
   ./segment -e ../models/yolo11n-seg.engine -i ../images -o ../output -l ../labels.txt --cudaGraph

Important

When inferring with an OBB model built using the --fp16 flag, there may be instances of duplicate anchor boxes. This issue is typically caused by a reduction in precision. Therefore, it is not recommended to build OBB models using the --fp16 precision mode.

Through the above methods, you can successfully complete model inference.