CTFS: Collaborative Teacher Framework for Forward-Looking Sonar Image Semantic Segmentation with Extremely Limited Labels

arXiv:2603.21071 · cs.CV, cs.AI · Submitted 2026-03-22 · Read on arXiv

Listen

Transcript

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

Tom: We're looking at the paper: "CTFS: Collaborative Teacher Framework for Forward-Looking Sonar Image Semantic Segmentation with Extremely Limited Labels".

Jane: The CTFS framework addresses performance limitations in semantic segmentation for forward-looking sonar images by introducing a collaborative teacher mechanism and a multi-view reliability assessment strategy to handle extremely limited labeled…

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

Paper summary: Tom: So Jane, so we’re looking at this paper called "CTFS: Collaborative Teacher Framework for Forward-Looking Sonar Image Semantic Segmentation with Extremely Limited Labels". What's the main idea here?

Jane: Well, it looks like the core problem they're tackling is that forward-looking sonar images have some really tough characteristics—like severe speckle noise, low texture contrast, and acoustic shadows—which makes traditional teacher student setups struggle when you only have a tiny amount of labeled data.

Lu: Exactly. The paper says these factors make generating good pseudo-labels from the teachers really difficult because the teacher models can produce a lot of low-quality predictions, sometimes worse than just using the ground truth baseline when you only use ten percent of your labeled data five twenty-four thirty-seven <ref:2603.21071#pg2>.

Meng: So what’s their proposed solution then? Are they just tweaking the existing setup or something new? I need to know if this actually helps us deploy anything.

Tom: They propose a whole new framework called CTFS. The thesis is that you need a multi-teacher collaborative mechanism, which means one general teacher and several sonar-specific teachers, instead of just relying on one type of teacher.

Jane: That multi-teacher approach uses something they call a multi-teacher alternating guidance strategy so the student model can learn general features while also picking up those specific sonar image traits at the same time.

Lu: The idea is to get more comprehensive and robust feature modeling by having different teachers provide different perspectives on the image data <ref:2603.21071#pg2>. They construct these teachers to simulate different types of noise and distortion, like Tgeneral using geometric transformations or color perturbations.

Meng: So what are the specific sonar-specific teachers they built? Are these just random noise models, or something tied directly to the physics of sonar imaging?

Tom: They have Tsonar a which focuses on simulating environmental interference factors like acoustic shadows, and Tsonar b which simulates acoustic imaging characteristics such as energy attenuation effects <ref:2603.21071#pg3>.

Jane: And that leads into their second big idea, which is this cross-teacher reliability assessment mechanism. They recognize that these teachers can generate a lot of noisy pseudo-labels, so they design something to dynamically quantify how reliable those labels actually are.

Lu: This reliability mechanism checks consistency and stability across multiple views and multiple teachers to mitigate the negative impact of those noisy pseudo-labels <ref:2603.21071#pg3>. They use single-teacher intrinsic stability verification and cross-teacher consistency verification to get a final reliability score for each pixel <ref:2603.21071#pg3>.

Paper summary: Meng: That sounds like a lot of computation just to check quality, which is what I worry about when scaling up. What happens when we have to run this assessment on every single unlabeled image?

Tom: They use that reliability score to define the unsupervised loss in their framework, called the Reliability-Guided Adaptive Constraint <ref:2603.21071#pg3>. This loss uses that reliability score along with a threshold of zero point four and a consistency smoothing factor of zero point five to filter out bad pseudo-labels or weight them softly <ref:2603.21071#pg3>.

Jane: So, what’s the actual result they found when they tested this on their benchmark data, the FLSMD dataset? That’s where we see if it actually works in practice.

Tom: On the FLSMD dataset, when only two percent of the data was labeled—which is extremely little—their method achieved a five point zero eight percent improvement in mIoU compared to other state-of-the-art approaches <ref:2603.21071#pg2>. That’s a significant bump just with minimal supervision.

Lu: It shows that when data is scarce, this framework actually performs better than what we see from other methods on the FLSMD dataset five twenty-four thirty-seven <ref:2603.21071#pg2>. The way they designed the collaboration between teachers seems to be where it gets its traction.

Meng: But I gotta ask about the complexity. They mention an overall framework in Figure two(a) with all these different teacher activations and assessment steps <ref:2603.21071#pg3>. How much extra training time does this add compared to just using a standard teacher student setup?

Tom: Well, they show that on the FLSMD dataset, when trained with DINOv2-S, their model achieves an mIoU of sixty-seven point zero seven when only two percent of the data is labeled <ref:2603.21071#pg3>. That’s a solid number for such low supervision levels.

Jane: It means that for someone who isn't reading the math, what this actually changes for them is that they can now get much better segmentation results even when they are severely limited in how much data they have to label.

Lu: The implication is that you don't have to rely on massive labeled datasets to get good sonar segmentation; you can use smarter, more specialized teacher guidance instead <ref:2603.21071#pg2>. This suggests that the way they structure the knowledge transfer is what matters most for capturing those unique sonar image properties.

Meng: For production, I’m still skeptical about how stable this reliability score stays when you move it to real traffic data, but seeing that performance improvement on FLSMD with only two percent labeled data is certainly something to consider.

Paper summary: Tom: So to wrap up the concept of CTFS: it’s a multi-teacher collaborative learning mechanism combined with a multi-view reliability assessment strategy specifically for forward-looking sonar images <ref:2603.21071#pg1>. It handles the limited data by using multiple perspectives and checking if those generated labels are trustworthy.

Jane: The authors, Ping Guo, Chengzhou Li, Guanchen Meng, Qi Jia, Jinyuan Liu Zhu Liu, Yu Liu, and Zhongxuan Luo Xin Fan from Dalian University of Technology put this together to tackle those challenging sonar image conditions.

Lu: Their work suggests that designing teachers that mimic the real distortions—like acoustic shadows or energy attenuation—is crucial because it lets the student model learn the actual physics of what it's looking at <ref:2603.21071#pg3>.

Meng: It seems like a sophisticated way to handle data scarcity, but I’d want to see how much computational overhead this actually adds when you try to deploy it on something that isn't the clean FLSMD benchmark.

Tom: The real excitement here is seeing that five point zero eight percent improvement in mIoU on the FLSMD dataset with only two percent labeled data <ref:2603.21071#pg2>. That number, for such a tiny labeled set, shows the mechanism is effective.

Jane: It really does show that even when you’re dealing with incredibly limited labels in a complex domain like sonar imaging, you can use structured collaboration to push performance up significantly.

Lu: The future work they hint at is likely focusing on how to make this framework even more adaptive or robust when the noise levels get worse than what they tested <ref:2603.21071#pg2>.

Meng: I hope future work addresses the deployment side, because right now, it’s hard to trust any method that relies on generating so many pseudo-labels without rigorous testing on actual operational data.

Tom: So, to summarize for listeners: CTFS is a framework that uses multiple specialized teachers—general and sonar specific—and checks the quality of those generated labels across different perspectives to get better segmentation even with very little labeled training data <ref:2603.21071#pg1>.

Jane: It means we can achieve better semantic understanding of underwater sonar images when we don't have thousands of perfect examples to train our models on.

Lu: The key idea is transferring knowledge in a way that respects the unique physical challenges inherent in the sonar data itself <ref:2603.21071#pg3>.

Meng: I just hope they figure out how to keep that reliability assessment manageable when we move past this benchmark and onto real-world, messy sonar imagery.

Tom: That’s the main challenge for any practical implementation of CTFS—making sure the assessment mechanism doesn't become a bottleneck during training.

Jane: And that's what we can look forward to seeing in the next steps of this research direction.

Conclusion: Tom: So we’ve been talking about CTFS and how it handles those really tough sonar images when you have almost no training data at all <ref:2603.21071#pg1>. Now we’re getting to the conclusion of this paper, so what does the title actually mean?

Jane: It’s "CTFS: Collaborative Teacher Framework for Forward-Looking Sonar Image Semantic Segmentation with Extremely Limited Labels" so it tells us it uses a team of teachers and a special way to segment those sonar pictures when we only have tiny amounts of labeled data.

Lu: The main implication is that the model isn't just relying on one teacher anymore, but this whole collaborative setup helps the student model learn both general stuff and those specific sonar image traits at the same time <ref:2603.21071#pg3>.

Meng: From a deployment standpoint, it means we can probably get decent segmentation results even when our labeled dataset is really small, which is huge for systems where data collection is expensive or slow.

Tom: Exactly. It moves the goalpost on how much labeled data you need to get usable results in these challenging domains.

Jane: And they show that this method actually improved the mIoU by five point eight percent on the FLSMD dataset when only two percent of its data was labeled <ref:2603.21071#pg2>. That’s a concrete number showing it works.

Lu: That five point zero eight percent improvement is significant because it beats other methods we've seen on that same data setup <ref:2603.21071#pg2>.

Meng: I still have to think about the computational cost of running all these different teachers and reliability checks, though. It sounds like a lot of extra overhead for inference when you’re trying to keep things fast at scale.

Tom: That is the engineering question, Meng. But the authors say that on their best results, setting a consistency smoothing factor delta to zero point five and a reliability score threshold psi to zero point four gave them the best performance <ref:2603.21071#pg3>.

Jane: So they found a sweet spot for how much trust the model should put in its generated labels, which is crucial when you don't have enough real examples to verify things manually.

Lu: It shows that tuning those parameters, like delta and psi, is key to making this framework work well across different types of data scarcity issues <ref:2603.21071#pg3>.

Tom: So the core message is that collaboration between general and specific teachers plus a reliability check saves the day when you have very little labeled sonar data.

Jane: It really suggests that for underwater sensing, we can stop just looking for huge datasets and start focusing on smarter ways to teach the model what sonar images actually look like.

Lu: The next thing they point toward is how this framework adapts when the noise levels in the real world are even worse than what they tested <ref:2603.21071#pg2>.

School of Software Technology, Dalian University of Technology

cs.CV, cs.AI

Submitted: 2026-03-22

Updated: 2026-10-19

Comments: Accepted to CVPR 2026 Finding. Code: https://github.com/pingggg516/CTFS

Code: https://github.com/pingggg516/CTFS

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

Importance score: 73/100

The gist: The CTFS framework addresses performance limitations in semantic segmentation for forward-looking sonar images by introducing a collaborative teacher mechanism and a multi-view reliability assessment

Key concepts

Multi-teacher collaborative mechanism: This involves using one general teacher and several sonar-specific teachers to guide the student model. The teachers alternate guidance, allowing the student to learn both broad semantic concepts and specific features unique to sonar images, leading to a more robust feature model.
Cross-teacher reliability assessment (MVRA): This mechanism dynamically checks if predicted labels are trustworthy by comparing results from different views and multiple teachers. It calculates stability scores based on prediction consistency across these sources, helping to filter out unreliable pseudo-labels generated during training.
Reliability-Guided Adaptive Constraint: This strategy uses the reliability score from MVRA to define the unsupervised loss for the student model. This allows the model to selectively weigh or filter pseudo-labels based on their calculated trustworthiness, ensuring that noisy labels do not negatively impact learning.

Terminology

Pseudo-labels — 3× in this paper · explained in 2 episodes
Semantic Segmentation — 3× in this paper · explained in 3 episodes
Consistency — 1× in this paper · explained in 3 episodes

Summary

The CTFS framework addresses performance limitations in semantic segmentation for forward-looking sonar images by introducing a collaborative teacher mechanism and a multi-view reliability assessment strategy to handle extremely limited labeled data. The gist: CTFS proposes a Collaborative Teacher Semantic Segmentation Framework for forward-looking sonar images that achieves a 5.08% improvement in mIoU on the FLSMD dataset when only 2% of the data is labeled, surpassing other state-of-the-art approaches

How it works

The framework introduces a multi-teacher collaborative mechanism composed of one general teacher and multiple sonar-specific teachers to transfer multi-perspective knowledge to the student model. This mechanism employs a multi-teacher alternating guidance strategy so that the student model can learn general semantic representations while simultaneously capturing the unique characteristics of sonar images, leading to more comprehensive and robust feature modeling.

The core training objective for the student model is defined by Equation (1), which is given as Ltotal = Lsup + λu · Lunsup. The supervised loss Lsup is calculated using the ground truth with a small amount of labeled data, and the unsupervised loss Lunsup will be discussed in the chapter on Reliability-Guided Adaptive Constraint.

Collaborative Teacher for Sonar (CBTS)

The CTFS method proposes multiple dedicated teacher models that are cyclically activated at different knowledge feature levels to provide multi-dimensional supervision and guidance to the student model. The three teacher models—Tgeneral, Tsonar a, and Tsonar b—are activated in a fixed order of ”general → sonar a → sonar b” to guide the same student model.

Different teachers activate with different perturbation combinations: Tgeneral uses traditional geometric transformations and color perturbations, while Tsonar a focuses on simulating environmental interference factors such as acoustic shadows, and Tsonar b focuses on simulating acoustic imaging characteristics such as energy attenuation effects.

The prediction results of the teacher models for unlabeled images are calculated by Equation (5), which is given as p t u = X t∈T Pϕ(e) A t w(Xu). The acoustic shadow perturbation process designed for Tsonar a is represented by Equation (6), which defines Io(x, y) = (Ii(x, y) × (1 − α / 1 − d/R), if (x, y) ∈ S, Ii(x, y), otherwise.

Multi-view Reliability Assessment (MVRA)

The Multi-view Reliability Assessment mechanism is designed to dynamically assess the reliability of pseudo-labels by evaluating the consistency and stability of predictions across multiple views and multiple teachers. The mechanism consists of two core parts: single-teacher intrinsic stability verification and cross-teacher consistency verification.

Single-teacher intrinsic stability verification calculates the prediction stability score r t ij using Equation (13), which is given as r t ij = 1 / NAtw A t Xw k=1 cosf o t ij, f kt ij. Cross-teacher consistency verification assesses reliability based on the consistency among outputs of three teachers for the same unlabeled image using Equation (14), which is given as Cij = 1 / ND X(p,q)∈D cosf op ij, f oq ij.

The final reliability score for each grid cell is obtained by fusing the intrinsic stability of each single teacher and the predictive consistency of multiple teachers using Equation (15), which is given as Rij = ΠCij × 1 / NT X t∈T r t ij, where ΠCij denotes the Penalty item based on consistency of multi-teacher predictions.

Reliability-Guided Adaptive Constraint

The multi-view reliability score is used to define the unsupervised loss in the CTFS framework by Equation (17), which is given as Lunsup = 1 / Nu X Nu n=1 PH×W b=1 LCE(p s n[b], p t n[b]) × Rn b × ∆ Np, where ∆ is a binary vector used to determine whether a pixel participates in the loss calculation.

The reliability score threshold ψ is set to 0.4 and the consistency smoothing factor δ is set to 0.5, which are found to yield the best performance when analyzing key parameters. This approach allows the model to apply a reliability threshold for hard filtering of high-quality pseudo-labels or use it as a soft weighting factor in the loss function.

Sonar Semantic Segmentation Benchmark

The study introduces the Forward-Looking Sonar Image Semantic Segmentation (FSSG) dataset, collected in Bohai Bay using a remotely operated vehicle (ROV) equipped with a multibeam forwardlooking sonar (Oculus M750d). This dataset contains 3761 images from 11 target categories.

Experiments were conducted on both the FLSMD and FSSG datasets, sampling labeled data at ratios of 2%, 5%, and 10% of the training set. The results demonstrate that CTFS consistently outperforms current state-of-the-art methods on mIoU.

Experiment Results

Table 1 shows the quantitative comparison of mIoU (%) on the FLSMD and FSSG datasets with labeled data of 2%, 5%, and 10%. CTFS achieves a mIoU of 67.07 when trained with DINOv2-S on the FLSMD dataset with only 2% labeled data.

Table 3 shows the ablation study on the effectiveness of sonar-specific perturbations designed for the two additional sonar teachers. The results show that both sonar-specific perturbations bring significant and stable improvements across different aspects.

The model achieves its best performance when the consistency smoothing factor δ is set to 0.5 and the reliability score threshold ψ is set to 0.4. The results show that a 32×32 grid outperforms other sizes.

Conclusion

The CTFS framework successfully introduces a collaborative teacher mechanism and a multi-view reliability assessment strategy to enhance semantic segmentation performance for sonar images with extremely limited labels. This framework demonstrates remarkable superiority over other methods with extremely limited labels. The framework allows the student model to better capture the distinctive characteristics of sonar images and effectively suppresses the influence of noisy pseudo-labels during training.

--- Page 1 ---

CTFS: Collaborative Teacher Framework for Forward-Looking Sonar Image

Semantic Segmentation with Extremely Limited Labels Ping Guo, Chengzhou Li, Guanchen Meng, Qi Jia, Jinyuan Liu Zhu Liu, Yu Liu, Zhongxuan Luo Xin Fan School of Software Technology, Dalian University of Technology, Dalian, China Abstract As one of the most important underwater sensing technologies as one of the most important underwater sensing technologies Ping Guo1, Chengzhou Li1, Guanchen Meng1, Qi Jia1*, Jinyuan Liu1 Zhu Liu1, Yu Liu1, Zhongxuan Luo1 Xin Fan 1School of Software Technology, Dalian University of Technology, Dalian, China Semantic Segmentation with Extremely Limited Labels Ping Guo, Chengzhou Li, Guanchen Meng, Qi Jia 1Jinyuan Liu

Zhu Liu

Yu Liu

Zhongxuan Luo Xin Fan Abstract As one of the most important underwater sensing technologies forward-looking sonar exhibits unique imaging characteristics Sonar images are often affected by severe speckle noise low texture contrast acoustic shadows and geometric distortions These factors make it difficult for traditional teacher–student frameworks to achieve satisfactory performance in sonar semantic segmentation tasks under extremely limited labeled data conditions To address this issue we propose a Collaborative Teacher Semantic Segmentation Framework for forward-looking sonar images This framework introduces a multi-teacher collaborative mechanism composed of one general teacher and multiple sonar-specific teachers By adopting a multi-teacher alternating guidance strategy the student model can learn general semantic representations while simultaneously capturing the unique characteristics of sonar images thereby achieving more comprehensive and robust feature modeling Considering the challenges of sonar images which can lead teachers to generate a large number of noisy pseudolabels we further design a cross-teacher reliability assessment mechanism This mechanism dynamically quantifies the reliability of pseudo-labels by evaluating the consistency and stability of predictions across multiple views and multiple teachers thereby mitigating the negative impact caused by noisy pseudo-labels Notably on the FLSMD dataset when only 2% of the data is labeled our method achieves a 5.08% improvement in mIoU compared to other state-of-the-art approaches Code: https://github.com/pingggg516/CTFS1. Introduction Seawater exhibits strong absorption of visible light and in turbid waters deep sea environments or nighttime conditions-1.45%! Query: Do we need a teacher-student framework for sonar image semantic segmentation with extremely limited labels Texture detail loss Speckle noise Low Resolution challenging! EMA Teacher Student 61.94% 60.

Improvements for AI systems

This system can achieve superior semantic segmentation on forward-looking sonar images with extremely limited labeled data by employing a multi-teacher collaborative mechanism composed of one general teacher and multiple sonar-specific teachers. This allows the student model to learn general semantic representations while simultaneously capturing the unique characteristics of sonar images, thereby achieving more comprehensive and robust feature modeling.
The system will mitigate the negative impact of noisy pseudo-labels by incorporating a cross-teacher reliability assessment mechanism which dynamically quantifies the reliability of pseudo-labels by evaluating the consistency and stability of predictions across multiple views and multiple teachers. This directly addresses the issue where the inherent characteristics of sonar images—such as sparse texture details, severe speckle noise, and low resolution—lead teachers to generate a large number of noisy pseudolabels.
The system can leverage domain-specific knowledge by utilizing different teacher models with specialized perturbations: Tsonar a focuses on simulating environmental interference factors such as acoustic shadows and Tsonar b focuses on simulating acoustic imaging characteristics such as energy attenuation effects. This ensures that the student model learns to better capture the unique characteristics of sonar imagery, improving its cross-domain adaptability.
The final unsupervised loss function is enhanced by incorporating a reliability score: Lunsup = 1/Nu Σ n=1 PH×W b=1 LCE(p s n [b], pt n [b]) × Rn b × ∆ Np, where Rn b denotes the reliability score of the b-th pixel in the n-th image and ∆ is a binary vector used to determine whether a pixel participates in the loss calculation. This allows for more fine-grained training guidance by using this score as a soft weighting factor in the loss function to provide more fine-grained training guidance.
The system can be tuned for optimal performance by analyzing experimental results, finding that the model achieves the best performance when the consistency smoothing factor δ is set to 0.5 and the reliability score threshold ψ is set to 0.4, which effectively filters out low-quality pseudo-labels and ensures the student model can continuously learn effective features from high quality pseudo-labels.