H2SuperEllipse: Superellipse Representation for Oriented Object Detection With Horizontal Box Supervision

Oriented object detection is important for aerial image interpretation, where objects often appear with arbitrary orientations and diverse shapes. However, obtaining supervision such as pixel-wise masks or rotated bounding boxes (RBoxes) is costly, motivating HBox-supervised methods such as H2RBox-v2. While using HBoxes reduces annotation costs, the assumption that an HBox is the external circumscribed box of an inferred RBox may not hold for round or near-round objects, and rectangular RBoxes can include substantial background for non-rectangular targets. This paper presents H2SuperEllipse, a contour-guided HBox-supervised detector that extends the five-parameter RBox to a six-parameter superellipse representation (xc, yc,w, h, θ, n) for both square-like and round-like objects. The shape exponent n describes instance-level object extent, while omitting n naturally recovers the standard RBox output. Without mask or shape-parameter annotations, Edge Refinement (ER) extracts edge cues inside HBoxes, Contour-Guided Superellipse Pseudo-Labeling (CGSP) estimates pseudo-labels for n, and Superellipse Circumscribed Loss supervises the predicted super-ellipse from HBox annotations. On DOTA-v1.0, H2SuperEllipse achieves 73.58% AP50, exceeding manual-prior and prior-free H2RBox-v2 by 1.42 and 4.06 points. On a matched iSAID subset, the superellipses improve C-IoU from 73.90% to 75.20% and reduce background inclusion from 19.60% to 9.00% relative to the RBoxes from the same predictions. These results suggest that the proposed superellipse representation offers a new balance between annotation efficiency and geometric accuracy.