YOLO26-OBB

This version of YOLOv26-OBB (Oriented Bounding Box) has been converted to run on the Axera NPU using w8a16 quantization. It is optimized for detecting rotated objects with high precision.

Compatible with Pulsar2 version: 5.1.

Convert tools links:

For those who are interested in model conversion, you can try to export axmodel through:

• The repo of AXera Platform, where you can get the detailed guide.
• Pulsar2 Link, How to Convert ONNX to axmodel

Support Platform

• AX650N/AX8850
  • M4N-Dock(爱芯派Pro)
  • M.2 Accelerator card
• AX630C
  • 爱芯派2
  • Module-LLM
  • LLM630 Compute Kit
• AX615
• AX637

Performance Statistics

AX650N

Model	Latency(ms) npu1	Latency(ms) npu2
yolo26n-obb	9.426	3.432
yolo26s-obb	24.456	9.612
yolo26m-obb	69.012	27.254
yolo26l-obb	88.472	36.599
yolo26x-obb	196.579	71.008

AX630C

Model	Latency(ms) npu1	Latency(ms) npu2
yolo26n-obb	29.515	23.724
yolo26s-obb	61.693	47.572
yolo26m-obb	161.482]	110.556
yolo26l-obb	205.981	148.211
yolo26x-obb	501.142	294.563

AX615

Model	Latency(ms) npu1
yolo26n-obb	54.514
yolo26s-obb	156.599

AX637

Model	Latency(ms)
yolo26s-obb	26.364
yolo26l-obb	88.019
yolo26x-obb	88.019

How to use

Download all files from this repository to the device.

python env requirement

pyaxengine

https://github.com/AXERA-TECH/pyaxengine

wget [https://github.com/AXERA-TECH/pyaxengine/releases/download/0.1.3.rc2/axengine-0.1.3-py3-none-any.whl](https://github.com/AXERA-TECH/pyaxengine/releases/download/0.1.3.rc2/axengine-0.1.3-py3-none-any.whl)
pip install axengine-0.1.3-py3-none-any.whl

Inference with AX650 Host, such as M4N-Dock(爱芯派Pro)

Input image:

run

 python3 ax_infer.py --model-path yolo26l-obb_npu3.axmodel --test-img boats.jpg

root@ax650:~/yolo26_obb# python3 ax_infer.py --model_path yolo26l-obb_npu3.axmodel --test_img boats.jpg
[INFO] Available providers:  ['AxEngineExecutionProvider', 'AXCLRTExecutionProvider']
usage: ax_infer.py [-h] [--model-path MODEL_PATH] [--test-img TEST_IMG] [--img-save-path IMG_SAVE_PATH]
                   [--score-thres SCORE_THRES] [--nms-thres NMS_THRES] [--num-classes NUM_CLASSES]
                   [--providers PROVIDERS]
ax_infer.py: error: unrecognized arguments: --model_path yolo26l-obb_npu3.axmodel --test_img boats.jpg
root@ax650:~/yolo26_obb# python3 ax_infer.py --model-path yolo26l-obb_npu3.axmodel --test-img boats.jpg
[INFO] Available providers:  ['AxEngineExecutionProvider', 'AXCLRTExecutionProvider']
[INFO] Using provider: AxEngineExecutionProvider
[INFO] Chip type: ChipType.MC50
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Engine version: 2.12.0s
[INFO] Model type: 2 (triple core)
[INFO] Compiler version: 5.2-dirty a6f1799a-dirty
[YOLO26-OBB] [13:21:27.889] [DEBUG] Load model time = 485.62 ms
[YOLO26-OBB] [13:21:27.934] [DEBUG] Pre-process time = 13.80 ms
[YOLO26-OBB] [13:21:27.990] [DEBUG] Forward time = 55.92 ms
[YOLO26-OBB] [13:21:28.004] [DEBUG] Post-process time = 12.79 ms
[YOLO26-OBB] [13:21:28.034] [INFO] Draw Results (182 oriented objects):
[YOLO26-OBB] [13:21:28.035] [INFO]   ship                 conf=0.90 cx=770.3 cy=183.5 w=103.2 h=31.2 theta=+23.6deg
...
[YOLO26-OBB] [13:21:28.144] [INFO]   ship                 conf=0.25 cx=484.9 cy=5.2 w=26.2 h=15.6 theta=+13.1deg
[YOLO26-OBB] [13:21:28.145] [INFO]   ship                 conf=0.25 cx=1770.5 cy=660.5 w=105.5 h=35.2 theta=+23.6deg
[YOLO26-OBB] [13:21:28.189] [INFO] Saved to result_yolo26_obb.jpg

Output image: