Region-Guided Search for Ultra-High-Resolution Small Object Detection

Abstract

Current object detection models face significant challenges when applied to ultra-high-resolution images, as global downscaling frequently destroys crucial details for small objects, while EGC wastes computation on irrelevant regions and causes object truncation at tile boundaries. We propose a two-step reasoning architecture that employs a Guide Model to identify semantically relevant spatial regions, followed by fine-grained detection on the selected high-resolution crops. We formalize EGC with a dynamic partitioning scheme to explicitly define its spatial coverage boundaries and establish an upper bound for spatial recall. Building on this formulation, we introduce a semantic alternative consisting of two Guide Model variants, Qwen-VL and YOLO+LLM, together with a multi-stage hybrid Detector Model integrating GroundingDINO, YOLOv8, and CLIP for verification. Experiments on 3,637 UHR images demonstrate that although grid cropping achieves the highest detection accuracy with an improvement of up to 26.9% mAP over the full-image baseline, our Guide Models recover most of this gain, improving mAP by 19.2% over the baseline while reducing the required number of crops by a factor of 3.7 and effectively mitigating false-positive over-detection.