yolo12 Target Detection

This test program will infer the yolo12 model to achieve target detection, and the results will be output in the form of printing and images.

Download the precompiled cvimodel

git clone https://github.com/Arielfoever/milkv-yolo12.git

Cross-compile YOLO program on PC

You can use the yolov8 program.

Duo256M yolov8 code location: sample_yolov8.cpp

You can also use yolov12 program which located at v12.cpp to get the output images and better printings.

Compilation method

Refer to the method in the previous chapter Introduction to compile the sample program.

After the compilation is completed, the sample_yolov8 program we need will be generated in the sample/cvi_yolo/ directory

Model compilation

If you have already downloaded the yolov8 repository, git pull is enough.

Export yolo11.onnx model

Download yolo12n.pt via YOLO12: Attention-Centric Object Detection

• Download the corresponding yolo11 model file

Take yolo11n as an example

After downloading, copy it to the ultralytics directory.

Install ultralytics.

Copy the yolo_export/yolov8_export.py code to the yolo11 repository, yolo11 is compatible with yolov8.

python3 yolov8_export.py --weights ./yolo12n.pt --img-size 640 640

TPU-MLIR conversion model

Please refer to the TPU-MLIR document to configure the TPU-MLIR working environment. For parameter analysis, please refer to the TPU-MLIR document.

After configuring the working environment, create a model_yolo11n directory in the same directory as this project and put the model and image files into it.

cp -rf ${REGRESSION_PATH}/dataset/COCO2017 .
cp -rf ${REGRESSION_PATH}/image .

onnx to MLIR

model_transform.py \
    --model_name yolo12n \
    --model_def ../yolo12n.onnx \
    --input_shapes [[1,3,640,640]] \
    --mean 0.0,0.0,0.0 \
    --scale 0.0039216,0.0039216,0.0039216 \
    --keep_aspect_ratio \
    --pixel_format rgb \
    --test_input ../image/dog.jpg \
    --test_result yolov12n_top_outputs.npz \
    --mlir yolov12n.mlir

After converting to mlir file, a yolo12n.mlir file will be generated.

MLIR to INT8 model (only supports INT8 quantized model)

Before quantizing to INT8 model, you need to run calibration.py to get the calibration table. The number of input data should be about 100~1000 according to the situation. Here, 100 COCO2017 pictures are prepared for demonstration:

run_calibration.py yolov12n.mlir \
--dataset ../COCO2017 \
--input_num 100 \
-o yolo12n_cali_table

Use calibration table to generate int8 symmetric cvimodel:

model_deploy.py \
--mlir yolov12n.mlir \
--quant_input --quant_output \
--quantize INT8 \
--calibration_table yolo12n_cali_table \
--processor cv181x \
--model yolo12n_cv181x_int8_sym.cvimodel

After compilation, a file named yolo11n_cv181x_int8_sym.cvimodel will be generated.

Board-side reasoning

Copy the compiled sample_yolov8, cvimodel, and the jpg image to be inferred to the board and execute the binary program:

scp sample_yolov8 yolo11n_cv181x_int8_sym.cvimodel 000000000632.jpg [email protected]:/root/

Execute the following command:

[root@milkv-duo]~# ./sample_yolov8 yolo12n_cv181x_int8_sym.cvimodel in.jpg  out.jpg
enter CVI_TDL_Get_YOLO_Preparam...
asign val 0
asign val 1
asign val 2
setup yolov8 param
enter CVI_TDL_Get_YOLO_Preparam...
setup yolov8 algorithm param
yolov8 algorithm parameters setup success!
---------------------openmodel-----------------------
version: 1.4.0
yolo12n Build at 2025-03-01 17:05:41 For platform cv181x
Max SharedMem size:6656000
---------------------to do detection-----------------------
image read,width:416
image read,hidth:640
objnum:6
Detect person(0): 143.069092 55.362289 340.261047 409.857300 0.854953
Detect person(0): 2.953743 35.646866 159.244690 479.729309 0.831542
Detect person(0): 342.131927 101.365387 415.000000 433.155334 0.743529
Detect dining table(60): 23.431549 390.899475 415.000000 639.000000 0.743529
Detect chair(56): 85.551483 371.947754 250.175781 448.465393 0.548328
Detect cake(55): 154.010742 430.204773 406.356934 562.873474 0.500000

See images at milkv-yolo12.