Sparse SAHI - how can it run?

[felix-br25]

First of all, this looks like a very promising project, kudos!!

One thing puzzles me in the Sparse SAHI example - how can it work softwarewise? I assume you have not modified the stgandard ultralytics code. So how can thsi example do soemthing new which is not part of the standard code?

from ultralytics import YOLO

model = YOLO("yolov8n.pt")

results = model.predict(
    source="large_aerial_image.jpg",
    sparse_sahi=True,
    slice_size=640,
    overlap_ratio=0.2,
    objectness_threshold=0.15,
)

I mean, it's the standard model, how can it run sparse_sahi?

I must be missing something.

[isLinXu]

Thanks for the feedback! Both issues have been fixed and merged into main.

Note: Sparse SAHI is still an experimental feature, and we've been actively optimizing the code recently. This commit addresses the two issues you raised and has passed testing.

Issue 1: Hard-coded objectness_threshold

Previously, the threshold was hard-coded to 0.15 at line 576 in predictor.py. This made the objectness_threshold=0.15 parameter in the README example effectively unusable, since the field wasn't defined in default.yaml.

Fix: We added objectness_threshold: 0.15 to default.yaml, and predictor.py now reads it dynamically via getattr(self.args, "objectness_threshold", 0.15). Users can now adjust the threshold freely.

Issue 2: Zero detections caused by global-model miss

The previous logic relied entirely on global detection results to generate the objectness mask. If the global model missed an object in a region, that region's objectness score became 0, causing all slices in that region to be skipped and resulting in zero detections.

Fix: We added the sparse_sahi_fallback parameter (default True). When the global model produces no detections, the pipeline automatically falls back to Full SAHI, running inference on all slices to prevent missed objects. If you need strict Sparse mode, you can disable it: sparse_sahi_fallback=False.

Usage

# Custom objectness threshold
results = model.predict(source, sparse_sahi=True, objectness_threshold=0.1)

# Disable fallback for strict Sparse mode (risk of missed detections)
results = model.predict(source, sparse_sahi=True, sparse_sahi_fallback=False)

Metadata enhancement

sparse_sahi_metadata now includes:

• fallback_slices: number of slices processed in fallback mode (useful for debugging)
• Source tagging: 0 = global, 1 = active slice, 2 = fallback slice

---

Please pull the latest main branch and give it a try. Sparse SAHI is still under experimental iteration, so feel free to open new issues if you run into anything else.