ViCrop-Det: Training-Free Small-Object Detection via Attention Entropy

A new method called ViCrop-Det uses spatial attention entropy to guide cropping for small-object detection without training. The approach addresses limitations of transformer-based architectures that apply uniform global receptive fields, causing feature degradation in dense microscopic target regions. ViCrop-Det is a training-free inference framework that adaptively shrinks spatial trust regions. It leverages the detection decoder's cross-attention distribution as an endogenous probe, evaluating local spatial ambiguity via Spatial Attention Entropy (SAE). This enables dynamic spatial routing, allocating a fixed computational budget exclusively to high-ambiguity zones. The method is inspired by attention entropy used in anomaly segmentation. The paper is published on arXiv with ID 2604.26806.