Object detection on resource-constrained platforms such as embedded and edge devices must operate under strict limitations in computational performance, memory capacity, and energy consumption. Conventional object detection pipelines process the entire image frame uniformly, regardless of the spatial distribution of relevant visual information. In many real-world scenarios, however, objects occupy only a small fraction of the image, enter and exit the field-of-view, or have been already detected. This leads to substantial computational effort being spent on regions without any useful information.

Inspired by biological visual attention mechanisms that prioritize salient regions of the visual field, we propose SIGHT, a saliency-driven selective tile processing framework for efficient object detection in resource-constrained environments. In SIGHT, each image frame is partitioned into tiles that are evaluated using a lightweight saliency estimation process, which assigns an importance score to each tile based on its visual relevance. Only tiles with sufficiently high saliency scores are forwarded to the object detection network (i.e. YOLO, tiny-YOLO, etc.), while uninformative regions are skipped.

By selectively processing only salient tiles, SIGHT significantly reduces the number of expensive inference operations required for object detection. Preliminary experimental results demonstrate that SIGHT achieves comparable detection accuracy to a full-frame YOLO-11 detector in identifying objects entering and leaving the observation scene, while reducing the total number of multiply–accumulate (MAC) operations by more than 50%. This reduction in computation directly translates into proportional savings in energy consumption, particularly important for battery-powered and embedded vision systems.