SAHI with Roboflow for Sliced Inference¶

In [1]:

#!pip install -U sahi
#!pip install -U inference rfdetr

#!pip install -U sahi #!pip install -U inference rfdetr

SAHI + Roboflow¶

This notebook demonstrates how to run object detection using Roboflow's RF-DETR models in combination with SAHI (Slicing Aided Hyper Inference) for improved detection performance—particularly on small objects.

You can set the model parameter in one of three ways, depending on whether you're using a model from Roboflow Universe or a custom-trained RF-DETR model.

Model Input Options¶

✅ Option 1: Roboflow Universe Model ID (as a string)¶

Use this when loading a model hosted on Roboflow Universe.

In [3]:

from sahi import AutoDetectionModel

detection_model = AutoDetectionModel.from_pretrained(
    model_type='roboflow',
    model='rfdetr-base',  # RoboFlow model ID as string
    confidence_threshold=0.5,
    device="cpu",
)

from sahi import AutoDetectionModel detection_model = AutoDetectionModel.from_pretrained( model_type='roboflow', model='rfdetr-base', # RoboFlow model ID as string confidence_threshold=0.5, device="cpu", )

✅ Option 2: Use a Model Class (e.g. RFDETRBase)¶

Use this when you want to explicitly define the model class without instantiating it.

In [4]:

from sahi import AutoDetectionModel
from rfdetr import RFDETRBase
from rfdetr.util.coco_classes import COCO_CLASSES

detection_model = AutoDetectionModel.from_pretrained(
    model_type='roboflow',
    model=RFDETRBase,  # pass the class itself
    confidence_threshold=0.5,
    category_mapping=COCO_CLASSES,
    device="cpu",
)

from sahi import AutoDetectionModel from rfdetr import RFDETRBase from rfdetr.util.coco_classes import COCO_CLASSES detection_model = AutoDetectionModel.from_pretrained( model_type='roboflow', model=RFDETRBase, # pass the class itself confidence_threshold=0.5, category_mapping=COCO_CLASSES, device="cpu", )

06/06/2025 21:59:52 - WARNING - rfdetr.main -   num_classes mismatch: pretrain weights has 90 classes, but your model has 80 classes
reinitializing detection head with 90 classes

Loading pretrain weights

✅ Option 3: Use a Model Instance¶

Use this if you have a pre-loaded or trained model instance.

In [5]:

from sahi import AutoDetectionModel
from rfdetr import RFDETRBase
from rfdetr.util.coco_classes import COCO_CLASSES

detection_model = AutoDetectionModel.from_pretrained(
    model_type='roboflow',
    model=RFDETRBase(),  # pass an instance of a trained model
    confidence_threshold=0.5,
    category_mapping=COCO_CLASSES,
    device="cpu",
)

from sahi import AutoDetectionModel from rfdetr import RFDETRBase from rfdetr.util.coco_classes import COCO_CLASSES detection_model = AutoDetectionModel.from_pretrained( model_type='roboflow', model=RFDETRBase(), # pass an instance of a trained model confidence_threshold=0.5, category_mapping=COCO_CLASSES, device="cpu", )

Loading pretrain weights

Running Inference¶

Once your detection_model is initialized using any of the above methods, you can perform inference using SAHI’s slicing and prediction API:

In [6]:

from sahi.utils.file import download_from_url

download_from_url('https://raw.githubusercontent.com/obss/sahi/main/demo/demo_data/small-vehicles1.jpeg', 'demo_data/small-vehicles1.jpeg')

from sahi.utils.file import download_from_url download_from_url('https://raw.githubusercontent.com/obss/sahi/main/demo/demo_data/small-vehicles1.jpeg', 'demo_data/small-vehicles1.jpeg')

In [7]:

from sahi.predict import get_prediction, get_sliced_prediction
from IPython.display import Image

result = get_prediction(
    "demo_data/small-vehicles1.jpeg",
    detection_model,
)
result.export_visuals(export_dir="demo_data/")
Image("demo_data/prediction_visual.png")

from sahi.predict import get_prediction, get_sliced_prediction from IPython.display import Image result = get_prediction( "demo_data/small-vehicles1.jpeg", detection_model, ) result.export_visuals(export_dir="demo_data/") Image("demo_data/prediction_visual.png")

UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:203.)
UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at /pytorch/aten/src/ATen/native/TensorShape.cpp:3637.)

Out[7]:

In [8]:

result_sliced = get_sliced_prediction(
    "demo_data/small-vehicles1.jpeg",
    detection_model,
    slice_height = 224,
    slice_width = 224,
    overlap_height_ratio = 0.2,
    overlap_width_ratio = 0.2,
)
result_sliced.export_visuals(export_dir="demo_data/")
Image("demo_data/prediction_visual.png")

result_sliced = get_sliced_prediction( "demo_data/small-vehicles1.jpeg", detection_model, slice_height = 224, slice_width = 224, overlap_height_ratio = 0.2, overlap_width_ratio = 0.2, ) result_sliced.export_visuals(export_dir="demo_data/") Image("demo_data/prediction_visual.png")

Performing prediction on 18 slices.

Out[8]:

In [ ]: