feat(detection): support windowed GeoTIFF reads in InferenceSlicer

[madhavcodez]

Description

Closes #2027.

InferenceSlicer can now run tiled inference directly on an open rasterio-style dataset, reading each tile via a windowed read instead of loading the whole image into memory. This makes it possible to slice multi-GB aerial/drone GeoTIFFs that don't fit in RAM — the exact pattern InferenceSlicer already implements, just sourced from disk windows rather than an in-memory array.

The original reporter outlined this with a rasterio.windows-based reference implementation; this PR integrates the capability into InferenceSlicer itself (the issue has been open and unclaimed for ~5 months).

import rasterio
import supervision as sv

def callback(tile):  # tile is (H, W, C); select/convert bands as needed
    return model.predict(tile)

slicer = sv.InferenceSlicer(callback, slice_wh=640, overlap_wh=100)

with rasterio.open("large_orthomosaic.tif") as dataset:
    detections = slicer(dataset)

Design

• rasterio stays fully optional — the core library imports zero rasterio symbols. Datasets are detected by duck typing (_is_windowed_raster: a callable .read plus .crs/.width/.height), and tiles are read with rasterio's tuple-window form read(window=((row_start, row_stop), (col_start, col_stop))). Installable via pip install "supervision[geotiff]" (new optional extra).
• Back-compat: the existing in-memory NumPy/PIL path is unchanged — same crop_image, resolution_wh, compact_masks, out-of-bounds warning, and move_detections semantics. All existing InferenceSlicer tests pass unmodified.
• CRS validation: a geographic (non-projected) CRS raises ValueError, since pixel-space tiling on lat/lon is meaningless; crs is None (a plain TIFF) is allowed.
• Tiles preserve the source dtype (e.g. uint16) and all bands; selecting/converting is the callback's responsibility (documented in the __call__ docstring).

Tests

tests/detection/tools/test_inference_slicer_geotiff.py — runs without rasterio installed via a lightweight fake dataset, and includes a real rasterio.MemoryFile integration test guarded by importorskip:

• windowed path produces the same merged detections as the in-memory array path (no overlap and with overlap_wh > 0);
• each tile handed to the callback is byte-identical to the crop_image tile;
• crs=None works, geographic CRS raises, projected CRS does not;
• source dtype (uint16) is preserved through to the callback.

Notes for reviewers

• Scope: this accepts an already-open dataset rather than a file path, deliberately, so that no rasterio import leaks into core. Happy to add a str | Path convenience overload (opening/closing the dataset internally) in a follow-up if you'd prefer that ergonomics.
• The accepted input type is expressed via a small structural WindowedRasterDataset Protocol so static type checkers accept a rasterio dataset without supervision depending on rasterio's types.
• Changelog updated under UnReleased.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 79%. Comparing base (918b613) to head (7354a40).

Additional details and impacted files

@@           Coverage Diff           @@
##           develop   #2281   +/-   ##
=======================================
  Coverage       79%     79%           
=======================================
  Files           66      66           
  Lines         8567    8584   +17     
=======================================
+ Hits          6726    6744   +18     
+ Misses        1841    1840    -1     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.