| ## Model Information | |
| ### Source Model | |
| - Input shape: 1024x1024 | |
| - Number of parameters: 23.49M | |
| - Model size: 48.36M | |
| - Output shape: 1x20x21504 | |
| - Source model repository: [YOLO26m-OBB](https://docs.ultralytics.com/tasks/obb/) | |
| ## Inference with AidLite SDK | |
| ```bash | |
| sudo aid-pkg update | |
| sudo aid-pkg install aidlite-sdk | |
| sudo aid-pkg install aidlite-qnn248 | |
| ``` | |
| ```bash | |
| cd run_yolo26m-obb | |
| python3.10 code/python/run_test.py \ | |
| --target_model ./models/QCS8550/FP16/yolo26m-obb_qcs8550_fp16.qnn248.ctx.bin \ | |
| --imgs ./code/python/boats.jpg \ | |
| --invoke_nums 10 | |
| ``` | |
| The converted package exposes `output0: [1, 20, 21504]`, i.e. `xywh` (4) + | |
| 15 DOTAv1 class scores + angle (1) for 21504 anchors. Decoding uses the | |
| Ultralytics OBB head layout and probabilistic-IoU NMS. | |
| ## Version | |
| `2026.11.05.1` | |