Tiling images at full resolution helps object detectors catch small objects

Standard detectors like YOLO are usually run on 640x640 inputs, so large photos are downscaled and tiny objects such as boats, people, and cars fall below the detection threshold. The proposed fix splits images into overlapping full-resolution tiles, runs the detector on each tile plus a downscaled whole-image pass, and maps boxes back to image coordinates. Duplicate detections from tile edges are merged using intersection-over-smaller-box matching alongside standard IoU. In tests on three public photos, tiling raised counts substantially, such as 5 to 91 boats in an aerial marina shot, at the cost of many more inferences. The author notes tiling can also add false positives and recommends density-estimation models for very dense crowds.