metadata
language:
- en
license: apache-2.0
tags:
- axera
- ax650
- axmodel
- object-detection
- visdrone
- yolo11
- pulsar2
datasets:
- Voxel51/VisDrone2019-DET
library_name: axengine
pipeline_tag: object-detection
VisDrone YOLO11s — AX650 Object Detection
YOLO11s trained on VisDrone2019-DET, compiled to AX650 AXMODEL via Pulsar2.
Model Details
| Item | Value |
|---|---|
| Architecture | YOLO11s |
| Task | Object Detection |
| Classes | 11 (pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, motor) |
| Input | 640×640 BGR, uint8 [0,255] → float [0,1] |
| Chip | AX650N (NPU3) |
| Quantization | INT8 (MinMax PerLayer) |
| Size | 10.2 MB |
| Latency | ~3.3 ms |
Repository Structure
models/ AXMODEL + model_meta.json
demo/ 5 VisDrone test images
python/ Python SDK (pydet + libdet.axera)
cpp/ Pre-compiled C++ binaries
bin/ visdrone_detect (aarch64)
lib/ libdet.so
include/ Headers
Usage
C++ (on AX650 board)
cd cpp
chmod +x bin/visdrone_detect
LD_LIBRARY_PATH=./lib:/soc/lib ./bin/visdrone_detect \
../models/model.axmodel ../demo/demo_00.jpg 0.25
Python (on AX650 board)
cd python
pip install -r requirements.txt
python example.py --model ../models/model.axmodel --image ../demo/demo_00.jpg
Requires pyaxengine on the target board.
Accuracy
| Output Layer | Cosine Similarity | MSE |
|---|---|---|
| output0 (80×80) | 0.99999 | 0.005 |
| output1 (40×40) | 1.00000 | 0.004 |
| output2 (20×20) | 0.99999 | 0.004 |
Preprocessing
Input uint8 BGR [0,255] is normalized to float [0,1] via std=1/255, matching the original ONNX model input range.
Limitations
- Fixed input: 640×640
- Batch size: 1
- Requires AX650 BSP SDK 3.10.2+ runtime