Looking for guidance: Implementing SAHI-like sliced inference for instance segmentation in C# (ONNX Runtime)

[cyh5757]

Hi,
I’m building an inference pipeline in C# using ONNX Runtime with an instance segmentation model exported to ONNX.

My dataset contains many small, densely packed objects, so running inference on the full image often increases FN (missed instances), and pushing higher resolution becomes expensive in terms of latency/memory. That’s why I’d like to implement SAHI-like sliced/tiling inference in C# (split image into overlapping tiles → run inference per tile → merge results back to the original image).

However, SAHI is Python-centric and I couldn’t find a solid C# reference, so I’m considering building a minimal SAHI-style C# implementation. Before I start, I’d appreciate guidance on best practices and implementation details:

Goals

• Split the input image into overlapping tiles

• Run instance segmentation inference per tile with ONNX Runtime

• Map detections/masks back to original coordinates

• Merge tile outputs (NMS / weighted merging / mask-based merging)

I’m not necessarily aiming for a full SAHI port at first—just a practical minimum viable sliced inference + merge that works reliably in production.

Thanks in advance!

[onuralpszr]

You can start what you said via using onnxruntime to build what you need and see how is it go, there different implementations of sahi so you won't be have much a trouble find resources or reference to begin with it.