PhysFlowNet: Learning Canonical Latent Manifolds via Spatio-Spectral Physics Priors for Underwater Object Detection¶
Conference: ECCV 2026
Paper: ECCV Official
Code: To be released
Area: Object Detection
Keywords: Underwater Object Detection, Physics Priors, Riemannian Manifold Rectification, Spatio-Spectral Embedding, Evidential Deep Learning
TL;DR¶
Addressing the non-linear feature manifold distortion caused by underwater light attenuation and scattering, PhysFlowNet models representation rectification as a physics-guided Riemannian feature retraction, coupling hierarchical spatio-spectral priors, structure-preserving downsampling, and evidential uncertainty calibration to suppress backscatter false alarms and establish new SOTA benchmarks.
Background & Motivation¶
Underwater optical imaging is fundamentally governed by wavelength-dependent absorption and particle scattering (formalized by the classical Jaffe–McGlamery formulation), which causes severe contrast attenuation, color distortion, and spatially non-uniform veiling light. In the latent representation space, these physical degradations severely distort the canonical feature manifold of clear scenes, breaking the Euclidean linear separability assumption that standard terrestrial detectors rely upon. Existing strategies exhibit critical drawbacks: cascaded Image Enhancement and Restoration (IER) pipelines impose prohibitive computational overhead and frequently over-smooth vital structural boundaries or hallucinate pseudo-artifacts that mislead downstream detectors; Domain Adaptation (DA) techniques struggle to capture the complex depth-varying heterogeneity across diverse water columns; standard CNNs without explicit physical modeling collapse into blind memorization, missing low-contrast targets while generating confident false positives driven by backscatter and marine snow.
The fundamental tension lies in treating physical degradation as an arbitrary black-box perturbation under an assumed Euclidean geometry. Real-world underwater degradation warps the underlying semantic manifold continuously and non-linearly. Furthermore, conventional strided downsampling acts as an additional cascaded low-pass filter, compounding the optical low-pass effects of turbidity and aggressively destroying high-frequency boundary and corner cues across scale transitions.
This paper tackles the challenge from a Riemannian flow perspective: feature rectification should be explicitly formalized as a geodesic transport on a physics-conditioned Riemannian manifold. Core Idea: extract joint spatio-spectral physical priors to dynamically condition a local Riemannian metric tensor and velocity field, executing a metric-preconditioned retraction that transports distorted features back to their canonical semantic manifold, complemented by peak-preserving downsampling and evidential uncertainty regulation.
Method¶
Overall Architecture¶
PhysFlowNet integrates physics-guided geometric rectification directly into an efficient single-stage detection backbone. The workflow begins by extracting global power spectrum statistics and local structural variations using the Hierarchical Physical Context Embedding (HPCE) to construct a unified physical condition \(z_{phy}\). In the backbone stages, the Parallel Physics-Residual Bottleneck (PPRB) applies a dual-stream design where a physics-conditioned branch estimates an SPD metric tensor and tangent velocity field to perform metric-preconditioned Riemannian retraction. Across spatial transitions, the Physics-Guided Manifold Downsampling (PMD) module retains subtle boundary responses by combining strided convolutions with peak-preserving max pooling. Finally, the network is trained with the Unified Evidential-Contrastive Objective (UECO), enforcing manifold compactness and Dirichlet-based epistemic uncertainty calibration.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
IN["Underwater Input<br/>RGB Image"] --> HPCE["Hierarchical Physical Context Embedding<br/>Global Spectral-Spatial Stats + Local Cues"]
IN --> BB["Backbone Semantic Stream<br/>Main Stream Features"]
HPCE -->|Physical Prior z_phy| PPRB["Parallel Physics-Residual Bottleneck<br/>Preconditioned Riemannian Retraction"]
BB --> PPRB
PPRB --> PMD["Physics-Guided Manifold Downsampling<br/>Geometric Path + Salient Peak Path"]
PMD --> HEAD["Detection Head & Regularization<br/>UECO (SCML Manifold Contrast + EUR Uncertainty)"]
Key Designs¶
1. Hierarchical Physical Context Embedding: Decoupled Global Spatio-Spectral and Local Structural Conditioning Because optical degradation exhibits both macro-scale turbidity shifts and micro-scale illumination variations, HPCE decomposes degradation estimation into two parallel paths. The global path computes the 2D Real Fast Fourier Transform (rFFT) power spectrum and aggregates channel-wise energy across \(K\) concentric frequency bands via learnable Softmax-normalized kernels, forming the spectral prior \(e_{freq} \in \mathbb{R}^{3K}\). Concurrently, channel-wise mean, standard deviation, and grayscale variance capture color cast and contrast attenuation as spatial statistics \(s_{spatial} \in \mathbb{R}^7\). These are gated and projected into a 128-D descriptor \(z_{global}\). In parallel, the local path processes the shallow feature map \(F_{main}\) using a \(5\times 5\) depthwise separable convolution to extract spatially varying context \(z_{local} \in \mathbb{R}^{B\times 64\times H'\times W'}\). Broadcasting and concatenating both yields the unified physical condition \(z_{phy} \in \mathbb{R}^{B\times 192\times H'\times W'}\), providing rich environmental parameters for subsequent manifold transport.
2. Parallel Physics-Residual Bottleneck: Metric-Preconditioned Riemannian Retraction Stream Standard Euclidean residual shortcuts cannot perform directional stretching or compression required to correct anisotropic feature distortion. PPRB decouples feature learning into a base semantic branch and an explicit geometric rectification branch. The velocity estimation sub-module modulates semantic features with \(z_{phy}\) via Adaptive Group Normalization (AdaGN) to predict the tangent update direction \(v\), bounded by a terminal \(\tanh\) activation to avoid early gradient divergence. The metric construction sub-module processes the concatenated features through a bottleneck to predict a diagonal Symmetric Positive Definite (SPD) tensor \(G_{diag} = \text{diag}(1 + \alpha \cdot \tanh(g))\), where \(\alpha=0.9\) caps the maximum metric distortion. Terminal convolutional weights are initialized to zero, ensuring that the metric tensor strictly begins as the Euclidean identity \(I\). The rectified feature is then computed via metric-preconditioned retraction: $\(F_{rec} = \text{GN}\left(F_{main} + v \odot (G_{diag} + \varepsilon)^{-1/2}\right)\)$ The displacement is injected through a learnable spatial gate \(G_{gate}\) and a zero-initialized channel scale \(\gamma\), allowing smooth transition from baseline Euclidean features to adaptive Riemannian rectification as training stabilizes.
3. Physics-Guided Manifold Downsampling: Counteracting Compounded Optical Low-Pass Smoothing Underwater forward scattering functions as a physical low-pass filter, and standard strided convolutions exacerbate this by locally averaging adjacent pixels, washing out fine-grained edges and small-scale targets across feature pyramids. PMD introduces a two-pathway scale transition: the geometric path uses standard strided convolution to maintain macro spatial semantics, while the physical path modulates input features with a degradation-aware gating network \(E_{phy}\) followed directly by max pooling: $\(y_{phy} = \text{Conv}_{1\times 1}\left(\text{MaxPool}\left(x \odot \left(1 + \sigma(E_{phy}(x))\right)\right)\right)\)$ Because max pooling propagates peak localized activations (such as corner and boundary responses) rather than corrupted neighborhood averages, structural cues that survived scattering are preserved across resolution transitions before being merged with the geometric path.
Loss & Training¶
The overall optimization objective combines standard detection losses with the proposed Unified Evidential-Contrastive Objective (UECO): $\(L_{total} = \lambda_1 L_{box} + \lambda_2 L_{dfl} + \lambda_3 L_{bce} + \lambda_4 L_{edl} + \lambda_5 L_{mani}\)$ where \(L_{box}\), \(L_{dfl}\), and \(L_{bce}\) denote CIoU bounding box regression, distribution focal loss, and binary cross-entropy classification losses, respectively.
UECO incorporates two complementary regularization mechanisms: - Supervised Contrastive Manifold Loss (SCML): Projecting foreground anchor features to the unit hypersphere \(S^{C-1}\), \(L_{mani}\) applies class-balanced supervised contrastive regularization across pyramid levels, pulling same-class representations closer along hyperspherical geodesics while pushing distinct classes apart. - Evidential Uncertainty Regulation (EUR): Formulated under Evidential Deep Learning (EDL), classification logits are mapped via clipped Softplus to Dirichlet evidence parameters \(\alpha_{i,c} = e_{i,c} + 1\). The risk loss \(L_{risk}\) minimizes prediction variance, while an annealed KL divergence regularizes non-target evidence: $\(L_{edl} = \frac{1}{|I|} \sum_{i\in I} \left[ L_{risk}^{(i)} + \min\left(1, \frac{t}{T_{anneal}}\right) \cdot \text{KL}\left(\text{Dir}(\tilde{\alpha}_i) \parallel \text{Dir}(\mathbf{1})\right) \right]\)$ This forces background scattering clutter and ambiguous floaters to exhibit high epistemic vacuity, effectively suppressing confident false-positive bounding boxes.
Key Experimental Results¶
Main Results¶
PhysFlowNet was extensively benchmarked on DUO, Brackish, TrashCan, and RUOD at \(640\times 640\) resolution, trained from scratch for 300 epochs on a single RTX 4090 GPU.
| Dataset | Method | AP | AP50 | AP75 | Params (M) | FLOPs (G) | FPS |
|---|---|---|---|---|---|---|---|
| DUO | RetinaNet | 57.8 | 78.9 | 63.0 | 55.38 | 83.2 | 16.2 |
| DUO | Faster R-CNN | 61.2 | 82.7 | 69.6 | 41.3 | 120.0 | 13.5 |
| DUO | Cascade R-CNN | 61.6 | 80.5 | 72.4 | 88.15 | 140.0 | 10.3 |
| DUO | YOLOv8 | 59.1 | 88.1 | 72.8 | 3.2 | 8.7 | 51.2 |
| DUO | YOLO11 (Baseline) | 60.8 | 89.2 | 73.9 | 2.6 | 6.7 | 55.7 |
| DUO | Boosting R-CNN | 63.7 | 79.0 | 72.3 | 45.95 | 169.0 | 22.0 |
| DUO | GCC-Net | 69.1 | 87.8 | 76.3 | 36.74 | 300.79 | 15.4 |
| DUO | WFDA | 66.3 | 90.4 | 79.4 | 2.6 | 7.2 | 54.0 |
| DUO | PhysFlowNet (Ours) | 69.3 | 92.7 | 80.4 | 8.19 | 14.0 | 39.7 |
Cross-benchmark performance comparison (Brackish, TrashCan, RUOD):
| Dataset | Metric | PhysFlowNet (Ours) | YOLO11 | WFDA (Prev. SOTA) | Gain over Baseline |
|---|---|---|---|---|---|
| Brackish | AP / AP50 | 86.1 / 99.1 | 80.4 / 98.1 | 82.6 / 98.5 | +5.7% / +1.0% |
| TrashCan | AP / AP50 | 49.2 / 65.4 | 45.1 / 61.9 | 46.9 / 63.9 | +4.1% / +3.5% |
| RUOD | AP / AP50 | 64.3 / 86.4 | 58.1 / 81.9 | 63.4 / 86.1 | +6.2% / +4.5% |
Ablation Study¶
Step-wise ablation on the TrashCan dataset under identical training configurations:
| Config | Geometry | Loss (\(L_{mani}\) / \(L_{edl}\)) | AP50 | AP | Note |
|---|---|---|---|---|---|
| Baseline Backbone (YOLO11) | Euclidean | - / - | 61.9 | 45.1 | Standard one-stage detector |
| + PMD | Euclidean | - / - | 63.0 | 46.4 | Preserves boundary extrema during downsampling (+1.3 AP) |
| + PMD + HPCE | Euclidean | - / - | 63.3 | 46.8 | Spatio-spectral physical conditioning (+0.4 AP) |
| + PMD + HPCE + PPRB (\(G=I\)) | Euclidean Residual | - / - | 63.6 | 47.2 | Tangent velocity updates without metric preconditioning (+0.4 AP) |
| + PMD + HPCE + PPRB (\(G_{diag}\)) | Riemannian Retraction | - / - | 64.9 | 48.4 | Diagonal SPD metric preconditioning yields major leap (+1.2 AP) |
| + \(L_{mani}\) | Riemannian Retraction | ✓ / - | 65.1 | 48.8 | Hyperspherical contrastive manifold regularization (+0.4 AP) |
| + \(L_{edl}\) (Full PhysFlowNet) | Riemannian Retraction | ✓ / ✓ | 65.4 | 49.2 | Dirichlet evidential uncertainty regulation (+0.4 AP) |
Key Findings¶
- Riemannian Metric Preconditioning Drives the Core Gain: Enabling the adaptive diagonal metric \(G_{diag}\) brings a major boost from 47.2 to 48.4 AP (+1.2%), significantly outperforming Euclidean AdaGN updates (+0.4%). This confirms that non-linear underwater degradation requires a Riemannian metric to locally stretch or contract distorted latent dimensions.
- PMD Mitigates Downsampling Information Collapse: Replacing standard strided convolutions with PMD improves baseline accuracy by +1.3% AP / +1.1% AP50. Visualizations show standard convolutions smudge low-contrast marine debris into the background, whereas PMD's max-pooling branch effectively retains boundary saliency.
- Evidential Modeling Filters Out Clutter: Multi-scale vacuity maps reveal that backscatter patches, ambient light flare, and marine snow exhibit high vacuity (epistemic uncertainty). EUR forces the network to penalize low-evidence false predictions, resolving the classic over-confidence failure of softmax detectors.
Highlights & Insights¶
- Grounded Physics-Geometry Synergy: Bridges the physical Jaffe–McGlamery model with Riemannian manifold flow, establishing an interpretable paradigm where degradation corresponds to feature drift and retraction transports features back to canonical coordinates.
- Zero-Initialized Metric Stability: The metric projection tensor is initialized to zero, ensuring \(G_{diag}\) starts strictly as the Euclidean identity \(I\). This eliminates early training divergence typically associated with Riemannian optimization on unconstrained manifolds.
- High Transferability: HPCE's decoupled frequency-spatial descriptor and PMD's peak-preserving scale reduction offer drop-in value for other scattering-dominated visual tasks, such as autonomous driving in fog/rain, medical endoscopy, and turbid aerial imaging.
Limitations & Future Work¶
- Diagonal Metric Tensor Simplification: To maintain 39.7 FPS real-time efficiency, the Riemannian metric is constrained to a diagonal tensor \(G_{diag}\), ignoring off-diagonal channel covariance curvature that could prove useful in extreme turbidity.
- Lack of Temporal Physical Priors: Underwater robotic operations typically process video streams where suspended matter displays continuous hydrodynamic motion; current single-frame HPCE does not leverage temporal coherence.
- Future Directions: Exploring low-rank off-diagonal SPD metric parameterizations to capture cross-channel curvature at bounded computational cost; extending Riemannian feature flows into spatiotemporal video domain representations.
Related Work & Insights¶
- vs. Cascaded IER Methods (e.g., Sea-thru, UIEC^2-Net): Two-stage enhancement-then-detection incurs excessive latency and frequently over-smooths subtle edges; PhysFlowNet performs direct latent feature rectification at 14.0 GFLOPs and 39.7 FPS.
- vs. Fixed Frequency Methods (e.g., WFDA): WFDA applies fixed wavelet decomposition, which cannot adapt to depth-dependent scattering variations; PhysFlowNet dynamically computes continuous spatio-spectral conditioning via learnable band weighting.
- vs. Domain Adaptation / Alignment (e.g., Boosting R-CNN, GCC-Net): Statistical alignment or adversarial discriminators focus on global domain shifts, neglecting spatially heterogeneous optical distortions; PhysFlowNet provides pixel-adaptive metric preconditioning.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ (Pioneering formulation of underwater feature rectification as physics-guided Riemannian flow with diagonal SPD metrics)
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ (Comprehensive validation across DUO, Brackish, TrashCan, and RUOD, with clean step-wise ablations isolating geometry, architecture, and loss regularizers)
- Writing Quality: ⭐⭐⭐⭐⭐ (Mathematically rigorous, well-structured, with clear correspondence between physical optics, geometric equations, and architectural modules)
- Value: ⭐⭐⭐⭐⭐ (Delivers both theoretical elegance and practical real-time performance for safety-critical underwater robotics and marine environmental monitoring)