Traditional object detection methods predominantly rely on bounding boxes for localization, often including substantial background regions depending on the object orientation and shape. This paper introduces a novel object detection paradigm termed “convex hull-ish object detection,” which employs convex hulls instead of bounding boxes for localization. This shift significantly mitigates background noise and can thus help extract better global features for characterizing detections. Our approach adopts a two-stage methodology. In the initial stage, we leverage an enhanced USegNet to generate approximate convex hulls, followed by a refining step utilizing a specialized convex hull algorithm. To validate our approach, we repurpose existing semantic segmentation datasets as convex hullish object detection benchmarks. Furthermore, a manually annotated dataset featuring hand-drawn convex hulls is presented to showcase an end-to-end framework for convex hullish object detection. Our extensive experiments demonstrate the efficacy of this new paradigm.