EVENT DOSSIER
The research team proposed the AquaBEV model, aimed at addressing the difficulty of directly predicting the occupancy of a local bird’s-eye view (BEV) from a single underwater RGB image. The model utilizes paired 3D imaging sonar as geometric supervision signals, maps visual features to uncalibrated polar coordinates, and reconstructs them into Cartesian BEV coordinates through causal decoding. To this end, the team established controlled underwater occupancy benchmarks to uniformly evaluate representative methods. Experimental results showed that AquaBEV achieved 31.4% and 38.6% respectively in the visible IoU and observed IoU metrics, achieving a relative improvement of 4.0% and 4.3% compared to the strongest migration baseline.
AquaBEV 模型利用配对 3D 成像声呐作为几何监督，从单张水下 RGB 图像预测局部鸟瞰图（BEV）占用。该模型将视觉特征映射至无校准极坐标表示，沿距离维度应用因果解码后重构为笛卡尔 BEV 坐标。研究团队建立了受控水下占用基准测试，统一适配代表性占用方法。AquaBEV 在可见 IoU 和观测 IoU 上分别达到 31.4% 和 38.6%，较最强迁移基线相对提升 4.0% 和 4.3%。