Arquimea's QueryMLP boosts buoy detection with explicit spatial priors

In the MaCVi 2026 Vision-to-Chart Data Association Challenge, participants must align buoy positions from nautical charts with their visual detections in camera images—a geometrically intricate task due to camera perspective, motion, and sea conditions. The challenge baseline used a DETR-based fusion transformer that received per-buoy queries encoding world-space distance and bearing, forcing the decoder to implicitly learn the full projection from 3D world coordinates to 2D image pixels. This implicit learning is computationally heavy and prone to errors when the geometric relationship is complex.

Researchers at Arquimea Research Center introduced a simple yet effective modification: QueryMLP, a dedicated multi-layer perceptron trained to explicitly predict each buoy's waterline contact point in pixel coordinates using chart measurements and onboard IMU orientation data. These predicted pixel coordinates were appended to the baseline decoder query vector, providing a direct spatial prior for each buoy. This offloaded the geometric reasoning from the transformer decoder, allowing it to focus on finer visual matching. On the held-out test set, the method scored an Overall metric of 0.7386 with F1=0.8055 and mIoU=0.6718, earning 2nd place among all submissions. The approach is lightweight (described in just 5 pages) and code is publicly available.