Trung Tien Dong
WACV 2027 · Submitted

AquaBEV: Monocular Underwater BEV Occupancy with 3D Sonar Supervision

Trung Tien Dong1, Shengji Jin2, Chen Chen3, Yi Sheng2, Xiaomin Lin1

1ERA Lab, University of South Florida  ·  2YES Lab, University of South Florida  ·  3University of Central Florida

arXiv Paper Code Coming soon
AquaBEV task formulation
Task. A single RGB image is used to predict local BEV occupancy. Paired 3D imaging sonar provides a three-state target (occupied / observed-free / unknown) during training only — at inference, AquaBEV needs the camera alone.

Abstract

Autonomous underwater robots are widely used for exploration, monitoring, and inspection, where safe navigation depends on understanding the surrounding free and occupied space. Bird's eye view (BEV) occupancy provides such a representation, but predicting it from a single underwater RGB image is difficult due to limited, unreliable geometric cues from appearance alone. 3D imaging sonar offers complementary geometric measurements to supervise this task.

We introduce AquaBEV, a monocular underwater occupancy model that predicts local BEV occupancy from a single RGB image, using paired 3D imaging sonar as geometric supervision during training. AquaBEV maps visual features into a calibration free polar representation and applies causal decoding along the range dimension before reconstructing the prediction in Cartesian BEV coordinates. A controlled underwater occupancy benchmark was established, adapting representative occupancy methods to the same RGB to sonar task under a unified protocol. AquaBEV achieves 31.4 Visible IoU and 38.6 Observed IoU, 4.0% and 4.3% relative improvements over the strongest transferred baseline.

31.4
Visible IoU
38.6
Observed IoU
+4.3%
relative Observed IoU over the strongest transferred baseline
RGB only
at inference — sonar is used for training supervision

Contributions

Method

AquaBEV architecture
Overview of AquaBEV. AquaBEV encodes a single RGB image into a learned polar representation (θ, r), applies causal reasoning along the range dimension, and differentiably transforms the resulting polar occupancy prediction into Cartesian BEV coordinates (x, y).

An image encoder extracts features that a calibration-free cross-view transformer gathers into a polar latent field indexed by azimuth and range, using learned polar queries. A causal range decoder then processes each ray from near to far without looking ahead — mirroring how an acoustic pulse propagates — before a differentiable warp reconstructs the prediction on a Cartesian BEV grid. Training supervises only the cells the sonar actually observed.

Results

Qualitative comparison
Qualitative comparison of monocular underwater BEV occupancy. From left: front-view RGB, sonar ground truth, AquaBEV (ours), MonoScene, SurroundOcc, TPVFormer and GaussianFormer.
Ablation
Ablation of the causal range decoder. Preserving the ordered range structure — not polar coordinates alone — is what drives the improvement.
Qualitative results with divers
More qualitative results with divers. Per row: RGB input, sonar ground truth, thresholded prediction, occupancy probability, and an error map (true positive / false positive / false negative).

BibTeX

@misc{dong2026aquabev,
  title         = {AquaBEV: Monocular Underwater BEV Occupancy with 3D Sonar Supervision},
  author        = {Trung Tien Dong and Shengji Jin and Chen Chen and Yi Sheng and Xiaomin Lin},
  year          = {2026},
  eprint        = {2609.04411},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2609.04411}
}