Skyline-Aware ROI-Calibrated Buoy Association for Vision-to-Chart Data Association

JO WONWOO, HyewonChun, Sangmunlee

Keywords: maritime computer vision, vision-to-chart association, buoy detection

Abstract:

The demand for reliable perception on unmanned surface
vehicles (USVs) is rapidly increasing, and the association
between onboard visual perception and chart data is essen-
tial for collision-free autonomous navigation. The problem
is particularly difficult on USVs, where aggressive steering
and acceleration induce large horizon changes while the
targets remain tiny, sparse, and often distant. The MaCVi
Vision-to-Chart Data Association challenge [7] targets this
exact problem. While end-to-end query-conditioned mod-
els offer an elegant unified formulation, they struggle to
converge reliably under the dynamic horizon variations in-
herent to USV operation. To overcome this, we propose
three key components: (i) explicit skyline detection and roll
compensation to stabilize the geometric reference frame, (ii)
chart-data-conditioned region-of-interest (ROI) projection
that constrains the feasible image region for each query, and
(iii) SAHI-based detector fine-tuning that sharpens localiza-
tion within the projected regions. As the winning solution to
the Vision-to-Chart Data Association challenge, our method
achieves an overall score of 0.7646 on the hidden test set.
The results demonstrate that horizon information is critical
for maritime perception systems, and that effectively integrat-
ing geometric priors into the association pipeline remains a
central challenge for robust maritime autonomy.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 15