Thermal-Adapted RF-DETR with Skyline-Guided Filtering and Cross-Scale Fusion for Maritime Thermal Object Detection

JO WONWOO, HyewonChun, Sangmunlee

Keywords: Thermal maritime detection, Object detection, Ensemble fusion, Auxiliary model agreement, Skyline-aware post-processing

Abstract:

Reliable thermal perception is essential for autonomous sur-
face vehicles, but maritime infrared detection remains dif-
ficult because targets are often small and low contrast, the
sea-sky boundary creates structured clutter, and most strong
detectors are pretrained on RGB imagery rather than single-
channel thermal data. The MaCVi 2026 thermal object
detection challenge [7] targets this setting. We address it
with four complementary components: (i) split refinement
that removes degraded frames and resolves an inconsistent
wind turbine annotation policy, (ii) a lightweight infrared
input adapter with staged encoder unfreezing for RF-DETR,
(iii) a cross-scale reranking module that estimates detection
quality from agreement across frozen prediction sources, and
(iv) three-checkpoint fusion with skyline-guided geometric
filtering to suppress above-horizon false positives. In the
MaCVi 2026 Thermal Object Detection Challenge, our final
system achieves 46.85 AP, placing third. These results show
that combining thermal-specific representation adaptation
with explicit maritime geometry priors is an effective strategy
for long-range infrared object detection.

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: 7