Source-linked AI summary
AquaBEV: Monocular Underwater BEV Occupancy with 3D Sonar Supervision
Trung Tien Dong, Shengji Jin, Chen Chen, Yi Sheng, Xiaomin Lin
TL;DR
Reliable underwater BEV occupancy is difficult to infer from a single RGB image because visual geometric cues are limited and unreliable. AquaBEV uses paired 3D imaging sonar for training supervision while predicting occupancy from RGB alone, with a learned polar representation and causal range decoding. It outperforms the strongest transferred baseline on Visible IoU and Observed IoU.
Problem
Underwater monocular BEV occupancy is difficult because visual appearance provides limited, unreliable geometric cues, while sonar offers complementary geometric measurements.
Source: §1 · p. 1Method
AquaBEV uses paired 3D imaging sonar as training supervision and predicts local Cartesian BEV occupancy from RGB alone through learned polar features and causal range decoding.
Source: §1 · pp. 1–2Results
AquaBEV outperforms the strongest transferred baseline by 4.0% in Visible IoU and 4.3% in Observed IoU.
Source: §1 · p. 2Takeaways & Limitations
Under the unified benchmark, AquaBEV achieves the strongest overall occupancy performance among evaluated methods while requiring only a single RGB image at inference.
Source: §6.1 · p. 8The study covers one underwater environment and evaluates AquaBEV only as a perception model; improved occupancy prediction does not yet establish improved navigation performance.
Source: §6.1 · p. 8
Abstract
from arXiv · showhide
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.
1. Introduction
Reliable underwater navigation requires distinguishing occupied from navigable space, but degraded visual cues make monocular BEV occupancy difficult. AquaBEV uses sonar supervision during training and predicts occupancy from RGB alone at inference.
1. Introduction
- Underwater occupancy is difficult because cameras face attenuation, scattering, illumination variation, and limited visibility, while acoustic observations are sparse and difficult to fuse with camera data.Occupancy maps based on monocular depth inherit limitations of the depth estimate, which degrades under changing visual conditions.Source: §1 · p. 1
- Sonar supplies geometric supervision during training, while the model predicts occupancy from RGB alone at inference, avoiding the added hardware and integration complexity of using both sensors at inference.The approach learns occupancy directly from visual observations without depth as an intermediate step.Source: §1 · p. 1
- Transferring terrestrial occupancy methods to underwater imagery, sparse acoustic targets, and uncertain sensor geometry remains unclear, motivating a controlled benchmark under a unified protocol.BEV occupancy organizes free, occupied, and unknown space in a ground-plane frame that downstream planning can consume directly.Source: §1 · pp. 1–2
- AquaBEV uses a learned polar representation and causal range decoding before differentiably reconstructing Cartesian BEV occupancy, reflecting sonar’s azimuth-and-range ordering.The paper presents this as the first study of monocular BEV occupancy in the underwater domain.Source: §1 · p. 2
- AquaBEV outperforms the strongest transferred baseline by 4.0% in Visible IoU and 4.3% in Observed IoU.The paper also establishes a controlled benchmark with representative occupancy methods under a unified training and evaluation protocol.Source: §1 · p. 2
2. Related Work
Related work connects underwater spatial mapping with BEV and occupancy methods, then reviews polar representations as a fit for radial sensing. Direct monocular underwater occupancy with sonar supervision remains distinct from these prior settings.
2.1. Underwater Navigation and Spatial Mapping
- Persistent underwater spatial representations support planning, but existing approaches generally rely on SLAM, geometric sensing, or estimated depth; dense occupancy directly from monocular underwater imagery remains largely unexplored.DREAM combines perception and spatial reasoning for monitoring [28], while CORAL uses occupancy for local planning and collision avoidance [29].Source: §2.1 · p. 2
2.2. BEV and Occupancy Prediction
- Terrestrial BEV and occupancy methods use explicit geometry, learned queries, dense or sparse representations, and video-based rendering supervision, but their transfer to degraded underwater imagery and sparse sonar supervision remains unclear.Cartesian representations are also poorly matched to supervision produced by a radial sensor.Source: §2.2 · p. 2
2.3. Polar and Range Structured Representations
- Polar and cylindrical methods preserve radial measurement structure, but prior approaches primarily process LiDAR directly; AquaBEV instead uses polar space as an intermediate for sonar-supervised occupancy prediction from perspective RGB.Prior work also addresses range-dependent distortions through feature correction, multiscale features, and range-aware processing.Source: §2.3 · p. 3
3. Methodology
AquaBEV learns local Cartesian BEV occupancy from RGB, using paired sonar returns to construct supervision during training. It organizes visual features by azimuth and range, causally decodes the range dimension, and transforms predictions into Cartesian coordinates for training.
3.1. Task Formulation and Sonar Supervision
- Paired sonar returns supervise RGB-based occupancy training; returns mark occupied cells, traversed rays mark observed free cells, and unobserved cells are excluded.Sonar is not provided to the network at inference.Source: §3.1 · p. 3, §3.6 · p. 4
3.2. AquaBEV Overview
- AquaBEV predicts Cartesian occupancy through a learned polar representation and avoids explicit camera-to-sonar extrinsic parameters, while remaining tied to the training sensor configuration.A visual encoder feeds polar queries, followed by causal decoding and a differentiable transformation to Cartesian BEV.Source: §3.2 · p. 3
3.3. Cross View Transformer
- Learned queries at each azimuth and range location refine image features into a dense polar field that is passed to the causal range decoder.The representation uses 96 azimuth bins and 96 range bins, spanning −48° to 48° and 0.4 m to 10 m.Source: §3.3 · p. 3
3.4. Causal Range Decoder
- The causal decoder makes each range output depend only on current and nearer features, using gated causal convolutions along range and neighboring-azimuth interactions.The four range-convolution dilation rates are 1, 2, 4, and 8.Source: §3.4 · p. 4
3.5. Polar to Cartesian Reconstruction
- Differentiable sampling transforms predicted polar logits into Cartesian BEV, allowing the Cartesian objective to train the polar decoder without independently rasterized polar labels.Gradients from the Cartesian objective propagate through the transformation into the polar decoder.Source: §3.5 · p. 4
3.6. Training Objective
- The complete model is trained with a Cartesian occupancy objective on acoustically observed cells; at inference it maps one RGB image directly to local Cartesian BEV occupancy.No polar labels or sonar measurements are network inputs.Source: §3.6 · p. 4
4. Experiments
Under a fixed RGB-to-sonar occupancy protocol, AquaBEV outperforms the strongest transferred baseline on the principal IoU metrics. Ablations distinguish a small gain from polar decoding alone from a larger reproducible gain with causal range decoding, while qualitative results identify conditions where predictions remain difficult.
4.1. Dataset and Evaluation Protocol
- The benchmark uses 110 scenes and approximately 96,000 paired observations, with a fixed session split and single-RGB-image inference for every method.Targets come from accumulated sonar and include occupied, observed free, and unknown cells; evaluation includes Visible and Observed IoU, forward-distance bands, boundary F1, and Chamfer distance.Source: §4.1 · pp. 4–5
4.2. Comparison with Existing Occupancy Methods
- 31.4 Visible IoU and 38.6 Observed IoU exceed GaussianFormer’s 30.2 and 37.0; AquaBEV leads Range Macro and has 0.340 m Chamfer distance.AquaBEV reaches 43.0 Near IoU, 42.1 Mid IoU, and 22.2 Far IoU; SurroundOcc has higher boundary F1 (48.4 versus 45.9).Source: §4.2 · p. 5
4.3. AquaBEV Ablation
- Polar decoding raises Visible IoU from 30.8 to 31.3, while causal range decoding raises Observed IoU from 37.5 to 38.6 (1.10 ± 0.20 points).Survival supervision, metric angular coupling, and full range context show no consistent improvement.Source: §4.3 · pp. 5–6, Table 2 · p. 5, Figure 4 · p. 7
4.4. Qualitative Results
- Qualitative comparisons show more continuous occupied regions and better-preserved extent, especially for elongated structures and separated regions, but visually indistinct or sparsely observed areas remain challenging.In such regions, AquaBEV may underestimate occupied extent, merge nearby structures, or miss distant occupancy.Source: §4.4 · p. 6, Figure 3 · p. 6
5. Discussion
Transferred occupancy methods show a narrow performance range, while ablations attribute AquaBEV’s main reproducible gain to causal range decoding rather than polar coordinates alone. The discussion cautions against strict sonar-return assumptions and favors range ordering over full propagation modeling.
5.1. Occupancy Transfer to Underwater Perception
- Across seven transferred methods in underwater monocular occupancy prediction, Visible IoU spans 28.7–30.2 and Observed IoU 34.2–37.0; this narrow range suggests representation changes alone have limited impact in this setting.Source: §5.1 · p. 7
- A possible bottleneck is reliable geometry recovery from one underwater image; learned feature transport offers an alternative when cross-sensor geometry is difficult to obtain.AquaBEV does not use camera and sonar extrinsic parameters during feature transport, but its improvement does not establish that explicit projection is generally inferior.Source: §5.1 · p. 7
- AquaBEV leads occupancy overlap and Chamfer distance, while SurroundOcc achieves higher boundary F1, so no single metric captures prediction quality.Overlap measures recovered occupied structure, boundary metrics emphasize contour alignment, and Chamfer measures spatial displacement; joint reporting matters for sparse underwater scenes.Source: §5.1 · p. 7
5.2. Why Causal Range Reasoning Helps
- Polar coordinates alone yield only a small improvement over the matched Cartesian decoder; the larger reproducible change appears with causal range decoding.This separates the benefit of range ordering from simply expressing the scene in polar coordinates.Source: §5.2 · p. 7
- AquaBEV lets nearby azimuth regions interact while ordering predictions along range, so each distance is predicted without information from farther locations.Source: §5.2 · p. 7
- Metric angular coupling does not improve prediction, and extending the causal receptive field across the full range sequence yields essentially the same Observed and Far IoU.The improvement therefore cannot be explained by greater context alone.Source: §5.2 · p. 8
- Explicit survival supervision gives no consistent gain, and the BEV target may place surfaces at different heights at multiple ranges along one direction.Thus, a simple free region followed by one acoustic return is not generally consistent with the supervision target.Source: §5.2 · p. 8
- The useful sonar-informed inductive bias is range ordering, not a complete physical model of sonar propagation.This exploits radial-sensing structure without requiring the occupancy target to satisfy an overly restrictive acoustic model.Source: §5.2 · p. 8
6. Conclusion
Under a unified protocol, AquaBEV achieves the strongest overall occupancy performance among evaluated methods while requiring only RGB at inference. The study is limited to one underwater environment and perception evaluation, with navigation tests and broader prediction tasks left for future work.
6.1. Summary of Findings
- AquaBEV achieves the strongest overall occupancy performance among evaluated methods while requiring only a single RGB image at inference.The study also established a controlled benchmark using offline 3D imaging sonar supervision.Source: §6.1 · p. 8
- Ablations find that causal range decoding produces the main reproducible gain, while survival supervision, range-conditioned angular coupling, and longer range context add no further improvement.The results favor ordered range reasoning over detailed modeling of sonar propagation.Source: §6.1 · p. 8
- The study covers one underwater environment and evaluates AquaBEV only as a perception model; improved occupancy prediction does not establish improved navigation performance.Sonar accumulation and projection also introduce geometric and temporal ambiguities into supervision.Source: §6.1 · p. 8
6.2. Future Work
- Future work will integrate AquaBEV with underwater navigation frameworks such as CORAL and use closed-loop experiments to assess path planning, collision avoidance, and navigation efficiency.Source: §6.2 · p. 8
- Further planned work extends binary occupancy to semantic BEV prediction and studies physics-informed generalization across water clarity, illumination, acoustic density, and sensor configuration changes.The aim is to examine transfer to environments and operating conditions not represented during training.Source: §6.2 · p. 8