Just change one place and it will be compatible with NPU RK3588 ???

[sunmouren]

First, modify the positions of the markers on the image,：

/home/orangepi/.local/lib/python3.8/site-packages/sahi/models/ultralytics.py

and then use the following code to make the prediction：

from sahi.models.ultralytics import UltralyticsDetectionModel
from sahi.predict import get_sliced_prediction

model = UltralyticsDetectionModel(
    model_path=f'{root}/yolo26{model_size}_rknn_model',
   confidence_threshold=0.25,
) 

results = get_sliced_prediction(main_img_path, model)

[onuralpszr]

@sunmouren can you check latest git code, I might be already unlock this, and soon we going to make a release so we can check this one out too.

cc @fcakyon

[lakshanthad]

@sunmouren I can confirm that with the latest SAHI and Ultralytics code, Rockchip RKNN models work!

I have verified this on a RK3588 and here are the results.

| Format       | Slice (s) | Prediction (s) | Postprocess (s) | Total (s) |
|--------------|-----------|----------------|-----------------|-----------|
| PyTorch      | 0.370654  | 2.717401       | 0.005007        | 3.093062  |
| RKNN FP16    | 0.059528  | 0.617929       | 0.004121        | 0.681578  |
| RKNN INT8    | 0.059605  | 0.429021       | 0.004546        | 0.493171  |

Please note that RKNN INT8 is used along with the below branch:
ultralytics/ultralytics#24703

Also, here is my script for reference:

from sahi.models.ultralytics import UltralyticsDetectionModel
from sahi.predict import get_sliced_prediction

model = UltralyticsDetectionModel(
    model_path="yolo26n_rknn_model",
    confidence_threshold=0.25,
)

main_img_path = "small-vehicles1.jpeg"

results = get_sliced_prediction(main_img_path, model)

results.export_visuals(export_dir="results")

print(results.durations_in_seconds)

cc: @onuralpszr @fcakyon