Tiling inference beats whole-image resizing for small objects in large images

On 2026-11-02, an article titled "How slice inference fixes small-object detection in large images" was published on DEV Community. Its author proposes replacing whole-image resizing with tiled slice inference as a way to address missed detections of small objects in large images.

The article notes that common real-time detectors, including YOLO and its relatives, are trained and run at around 640×640 pixels. A 3613×2400 aerial photo therefore gets shrunk by a factor of roughly 5.6 to fit the model input. Under this compression, a boat that is about 30 pixels long ends up at roughly 5 pixels, below the threshold at which the network can reliably identify it. Using a drone photo of a marina as an example, the author reports that pointing an off-the-shelf object detector directly at the image yields only five boats, even though dozens are actually present. In the author's framing, the model itself is not the weak link; the preprocessing pipeline discards the small objects before they ever reach the network.

To address this, the author substitutes tiled slicing for global resizing and compares the approach against a single full-image pass on the same YOLOX-S model. The trade-off is higher compute: the marina image required 35 slices. Based on the material available, the work remains at the method-demonstrate-and-compare stage, with no further validation or deployment follow-up reported yet.