BenthicDINO: Physics-Informed Self-Distillation for View-Invariant Side-Scan Sonar Representations

arXiv:2608.23215 · cs.CV, cs.AI, cs.LG · Submitted 2026-08-24 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "BenthicDINO: Physics-Informed Self-Distillation for View-Invariant Side-Scan Sonar Representations".

Jane: The gist:

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, looking at the title, BenthicDINO: Physics-Informed Self-Distillation for View-Invariant Side-Scan Sonar Representations, what’s the big picture here? It seems like they are trying to solve a very specific problem in underwater perception.

Jane: They are taking something that is inherently flawed—the way intensities mix physical reflectivity with acquisition geometry—and they are building a system that forces the learned features to ignore those geometric factors.

Lu: The key idea is using physics constraints and statistical independence, specifically the Hilbert-Schmidt Independence Criterion penalty, to ensure that the representation doesn't care if you look at the data from a slant range of one angle or another.

Meng: For someone who only listens to the show, what does this actually change for your daily work? Does this mean I can stop spending weeks labeling every single piece of sonar imagery?

Tom: It means that for tasks like mapping benthic habitats, you could achieve very high accuracy with just a small fraction of the labeled data because the model learns those underlying physical relationships on its own.

Jane: So it shifts the focus from manually labeling every single point to training a self-supervised framework that captures the physics of how sonar data is created.

Lu: The implication for future work seems to be pushing this toward real-world deployment on larger, more complex datasets, which is what they plan to do by evaluating it on things like AI4Shipwrecks.

Meng: I just want to know what the immediate next step for this model would be if we wanted to get a prototype running right away.

Tom: The next step is likely taking this pre-trained representation and fine-tuning it with very little labeled data for specific, high-value tasks, which they suggest will be done by using a lightweight multi-layer perceptron or a dedicated segmentation decoder.

Conclusion: Tom: So, BenthicDINO is this new framework they’ve put out that uses physics to stop sonar images from being fooled by how you look at them.

Jane: Right, it takes side-scan sonar imagery and makes sure the features are actually about the seabed, not just about the angle of acquisition.

Tom: The authors are using a self-supervised learning approach built on DINOv3, and they're enforcing these physical constraints to decouple the viewing dynamics from what’s actually underwater.

Jane: It seems like they’ve set up some specific training rules—like simulating speckle noise or gain miscalibration—and then adding a penalty that forces the features to be independent of those physical viewing parameters.

Tom: I like that they use the Hilbert-Schmidt Independence Criterion, which is a statistical way to measure how dependent the learned features are on things like slant range and incidence angle.

Jane: So, what does this mean for us who actually need to map these areas? It suggests we can get really good representations of benthic habitats without needing massive amounts of labeled data for every single piece of imagery.

Tom: Exactly, and they show that it achieves near-peak performance in mapping with just ten percent of the available labels, which is a big deal for data efficiency.

Jane: But they also showed strong generalization across different domains, meaning it should work well on new types of sonar data too, even if we haven't seen those exact scenarios before.

Tom: I’m looking forward to seeing how well this performs when we move it from training to actual deployment, because getting a model that runs fast and doesn't break at scale is the real challenge.

Jane: Right, so after seeing these results on the S3Seg dataset, what we need to look at next is how robust these features are when applied to more complex, real-world sonar scenarios.

Taqi Hamoda, Hayat Rajani, Nuno Gracias

Computer Vision and Robotics Research Institute of the University of Girona, Spain

cs.CV, cs.AI, cs.LG

Submitted: 2026-08-24

Updated: 2026-10-16

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: The gist: This work introduces a physics-informed self-distillation framework built upon DINOv3 to create view-invariant representations for side-scan sonar imagery by enforcing physical constraints

Key concepts

View-Invariance: This refers to making the learned features independent of how the sonar image was acquired. Traditional methods fail because different acquisition angles (like slant range or incidence angle) change the recorded intensity. BenthicDINO forces the model to ignore these viewing parameters, ensuring that features representing a specific seabed are consistent regardless of the camera's physical position during capture.
Physics-Guided Augmentations: These are artificial data distortions introduced during training that mimic real-world sonar noise and artifacts. They include simulating speckle noise, attenuation changes based on distance, and errors from gain miscalibration. By training the model on these physically realistic variations, it learns to filter out these acquisition-specific nuisances.
Hilbert-Schmidt Independence Criterion (HSIC): HSIC is a mathematical tool used to explicitly enforce statistical independence between the learned dense patch features and the physical viewing parameters. It acts as a penalty during training, ensuring that the model's internal feature representations do not statistically correlate with variables like incidence angle or slant range, directly achieving view-invariance.

Terminology

Distillation — 1× in this paper · explained in 2 episodes
Self-Distillation — 1× in this paper · explained in 7 episodes

Summary

The gist: This work introduces a physics-informed self-distillation framework built upon DINOv3 to create view-invariant representations for side-scan sonar imagery by enforcing physical constraints and statistical independence between features and viewing parameters.

How it works

The proposed methodology enforces view-invariance through two primary mechanisms: physically motivated augmentations that simulate speckle noise, rangedependent attenuation, and radiometric miscalibration; and a Hilbert-Schmidt Independence Criterion (HSIC) penalty that explicitly decouples learned dense patch features from physical viewing parameters<ref:2608.23215#pg2>. Furthermore, a dense, hierarchical feature fusion strategy across all four network stages is used to preserve fine-grained sediment details alongside deep semantic abstractions<ref:2608.23215#pg5>.

The framework operates on two complementary fronts during training<ref:2608.23215#pg4>. First, physics-guided augmentations reproduce dominant SSS nuisance factors during training: bounded additive Gaussian perturbations act as a denoising regularizer against speckle, linear synthetic Time-Varying Gain (TVG) profiles enforce invariance to uncompensated attenuation, and radiometric jitter simulates gain miscalibration<ref:2608.23215#pg6>. Second, a Hilbert-Schmidt Independence Criterion (HSIC) penalty explicitly minimizes the statistical dependence between the learned dense patch features and the measured viewing parameters<ref:2608.23215#pg7>.

DINOv3 Self-Distillation Framework

The base model is built on DINOv3, employing a ConvNeXt-v2-Tiny backbone to maximize data efficiency<ref:2608.23215#pg5>. The overall pre-training objective is a composite loss function designed to capture both global semantic understanding and fine-grained local structures: LP re = LDINO + LiBOT + 0.5 · LGram + 0.1 · LKoLeo<ref:2608.23215#pg4>. The Global Discriminative Loss (LDINO) enforces view-invariance at the global image level by aligning representations across different augmented views of the SSS input<ref:2608.23215#pg5>.

The Masked Latent Reconstruction loss (LiBOT) is used to instill dense prediction capabilities and enhance the model’s robustness to acoustic shadows<ref:2608.23215#pg4>. Consistent with the global DINO objective, Sinkhorn-Knopp centering is also applied to these patch-level projections<ref:2608.23215#pg5>. Feature Regularisation (LKoLeo) is applied on the dense embeddings of the student patches to prevent feature collapse and encourage a uniform distribution across the hypersphere<ref:2608.23215#pg4>. Gram Anchoring (LGram) stabilizes dense features, accelerates iBOT convergence, and improves performance on dense SSS tasks such as feature matching and segmentation<ref:2608.23215#pg5>.

ConvNeXt-v2 Backbone

The framework opts for a convolutional architecture to maximize data efficiency because ViTs lack the strong inductive biases inherent to CNNs, and the tiny variant prevents severe overfitting on limited acoustic datasets<ref:2608.23215#pg5>. To mitigate representation collapse when applying mask-based self-supervised learning, ConvNeXt-v2 is utilized because it introduces Global Response Normalization (GRN)<ref:2608.23215#pg5>. This calibration successfully maintains feature diversity and prevents channel saturation during the masked latent reconstruction of sonar patches<ref:2608.23215#pg5>.

The model extracts and aggregates hierarchical representations from all four stages of the network<ref:2608.23215#pg5>. Feature maps from stages 2, 3, and 4 are upsampled via bilinear interpolation to match the spatial resolution of stage 1<ref:2608.23215#pg5>. These upsampled feature maps are then concatenated along the channel dimension<ref:2608.23215#pg5>. A 3-layer Multi-Layer Perceptron (MLP) compresses the concatenated features into a unified 768-dimensional descriptor<ref:2608.23215#pg5>. This dense feature fusion effectively preserves the finegrained, high-frequency sediment details captured in the initial stage, while seamlessly integrating the deep, robust semantic abstraction generated in the deeper stages<ref:2608.23215#pg5>.

Enforcing View-Invariance

To remove residual geometric information that augmentations alone cannot suppress, statistical independence is enforced using the Hilbert-Schmidt Independence Criterion (HSIC)<ref:2608.23215#pg7>. The objective is to minimize the statistical dependence between the patch features Z and the viewing geometry V<ref:2608.23215#pg8>. This term penalizes the network whenever its feature distribution covaries with the incidence angle or the slant range<ref:2608.23215#pg8>.

To maintain training efficiency, Random Fourier Features (RFF) are utilized to approximate a Gaussian Radial Basis Function (RBF) kernel for the HSIC regularizer<ref:2608.23215#pg8>. This reduces the computational complexity of the HSIC regularizer to O(B), scaling linearly with the batch size<ref:2608.23215#pg8>.

Unsupervised Feature Space Evaluation

Unsupervised analysis of the fused, dense feature space (R768) reveals that a critical failure mode like dimensional collapse is avoided, as the effective rank was measured at 53.81<ref:2608.23215#pg8>. The intrinsic dimension of the data manifold was estimated to be 8.80, suggesting that local semantic neighborhoods can be defined by approximately 9 degrees of freedom<ref:2608.23215#pg8>.

Clustering analysis using HDBSCAN identified 6 distinct, highly dense clusters with an impressive filtered Silhouette score of 0.5359<ref:2608.23215#pg8>. This demonstrates that the self-distillation objective, coupled with our physics-guided augmentations, forces the representations into highly stable, well-separated density basins<ref:2608.23215#pg8>.

Supervised Downstream Evaluation on S3Seg

The fused representation strictly dominated individual stage outputs, peaking at 71.4% mIoU and 86.5% accuracy at full data scale<ref:2608.23215#pg9>. The framework exhibited exceptional data efficiency: utilizing just 10% of the labeled data, the model achieved 96% of its absolute peak performance<ref:2608.23215#pg9>.

The Linear Probe proved vastly superior to K-NN for navigating this specific embedding space, as K-NN heavily overfits and collapses on the Test set<ref:2608.23215#pg10>. The results confirm that selfsupervised pre-training can effectively eliminate the massive annotation bottleneck traditionally associated with SSS habitat mapping<ref:2608.23215#pg10>.

The work introduces a novel, physics-informed selfdistillation framework tailored for Side-Scan Sonar (SSS) imagery<ref:2608.23215#pg2>. It demonstrates that selfsupervised pre-training can effectively eliminate the massive annotation bottleneck traditionally associated with SSS habitat mapping<ref:2608.23215#pg10>. The framework exhibits exceptional data efficiency, achieving 96% of its absolute peak performance using only 10% of the available annotated data<ref:2608.23215#pg9>. The final fused representation reaches a mean Intersection over Union (mIoU) of 71.4% and an overall accuracy of 86.5% during supervised downstream tasks on the S3Seg dataset<ref:2608.23215#pg9>. The researchers plan to evaluate the model on publicly available datasets, such as the AI4Shipwrecks dataset<ref:2608.23215#pg10>. The current ablation studies utilized Linear Probing and K-Nearest Neighbors to strictly measure the quality of the frozen latent space<ref:2608.23215#pg10>. The final segmentation performance on the S3Seg dataset will be further elevated by utilizing a lightweight multi-layer perceptron (MLP) or a dedicated segmentation decoder<ref:2608.23215#pg10>. The work was supported by Spanish Government through the project ”Automated Seabed Analysis through SelfSupervised Deep Learning Sonar Technology (ASSiST)” under grant PID2023-149413OB-I00<ref:2608.23215#pg10>. The references include [1] C. Lei, H. Rajani, N. Gracias, R.

Improvements for AI systems

The improved AI system can perform robust, cross-view correspondence for side-scan sonar (SSS) imagery by achieving a view-invariant representation—one that responds to the structural seabed and not the sensor geometry through two primary mechanisms: physically motivated augmentations that simulate speckle noise, rangedependent attenuation, and radiometric miscalibration; and a Hilbert-Schmidt Independence Criterion (HSIC) penalty.
The system can achieve high data efficiency in downstream supervised tasks by utilizing only 10% of the available annotated data to reach a mean Intersection over Union (mIoU) of 71.4%, demonstrating that self-supervised pre-training can effectively eliminate the massive annotation bottleneck traditionally associated with SSS habitat mapping.
The improved system can generate stable, noise-free semantic clusters for benthic topographies by utilizing a dense, hierarchical feature fusion strategy across all four network stages to preserve fine-grained sediment details alongside deep semantic abstractions, which results in features that are natively well-separated in the latent space with a 0.0% noise ratio.