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

cs.CV cs.AI Ping Guo, Chengzhou Li, Guanchen Meng, Qi Jia, Jinyuan Liu, Zhu Liu, Yu Liu, Zhongxuan Luo, Xin Fan · Mar 22, 2026
Local to this browser
What it does
Forward-looking sonar images suffer from severe speckle noise, acoustic shadows, and energy attenuation that break standard semi-supervised teacher-student frameworks. This paper proposes CTFS, a collaborative multi-teacher architecture...
Why it matters
A cross-teacher reliability assessment mechanism filters noisy pseudo-labels by measuring prediction consistency across teacher views. The work matters because sonar annotation is expensive and existing methods fail with <10% labels due to...
Main concern
The paper presents a plausible domain-specific extension of semi-supervised semantic segmentation for sonar imagery, with physically-motivated augmentations and a sensible reliability mechanism. However, it suffers from an unverifiable...
Community signal
0
0 up · 0 down
Sign in to vote with arrows
AI Review AI reviewed
Plain-language introduction

Forward-looking sonar images suffer from severe speckle noise, acoustic shadows, and energy attenuation that break standard semi-supervised teacher-student frameworks. This paper proposes CTFS, a collaborative multi-teacher architecture where one general teacher and two sonar-specific teachers (simulating acoustic shadows and energy decay) alternate to guide a student model. A cross-teacher reliability assessment mechanism filters noisy pseudo-labels by measuring prediction consistency across teacher views. The work matters because sonar annotation is expensive and existing methods fail with <10% labels due to domain mismatch.

Critical review
Verdict
Bottom line

The paper presents a plausible domain-specific extension of semi-supervised semantic segmentation for sonar imagery, with physically-motivated augmentations and a sensible reliability mechanism. However, it suffers from an unverifiable 'first' claim, mathematical notation errors, inconsistent reporting of improvement margins (5.08% in abstract vs. 4.76% in Section 5.2.2), and incomplete experimental details. While the core ideas hold merit, the rushed presentation and lack of statistical rigor weaken the scientific contribution.

“To the best of our knowledge, the CTFS framework proposed in this paper is the first semi-supervised semantic segmentation framework specifically designed for the forward-looking sonar image.”
paper · Abstract
What holds up

The collaborative teacher design is well-motivated: the sonar-specific teachers employ physics-inspired perturbations—acoustic shadow simulation $I_o(x,y) = I_i(x,y) \times [1 - \alpha(1 - d/R)]$ for shadow regions $\mathcal{S}$ and directional energy attenuation $I_o(x,y) = I_i(x,y) \times (1 - \gamma \times y/H)$—that directly address imaging artifacts described in Figure 3. The Multi-view Reliability Assessment (MVRA) intelligently combines single-teacher stability (augmentation consistency) with cross-teacher consensus $C_{ij} = \frac{1}{N_{\mathcal{D}}} \sum_{(p,q) \in \mathcal{D}} \cos(f_{ij}^{op}, f_{ij}^{oq})$ to weight unsupervised loss, which is more robust than thresholding alone. Ablation studies (Table 2) validate incremental gains: baseline 51.08% → EMA 55.94% → +CBTS 59.94% → +MVRA 62.32% mIoU.

“This design is intended to simulate the shadow formed when sonar is blocked by an obstruction during propagation... This design aims to simulate the directional energy attenuation caused by factors such as seawater absorption of sound wave beams.”
paper · Section 3.2, Eq. 6 and 11
“M1: 51.08, M2: 55.94, M3: 59.94, M4: 62.32”
paper · Table 2
Main concerns

First, the claim of being the 'first' semi-supervised sonar segmentation framework is strong and likely false given existing literature (e.g., Li & Zhang 2024 on underwater waste segmentation with sonar images). Second, mathematical presentation is sloppy: Equation 5 confusingly sums over teachers $t \in T$ while $\phi(e)$ selects only one teacher per epoch; Equation 17 is truncated with '$\times\Delta\$'. Third, result reporting is inconsistent: the abstract claims a 5.08% mIoU improvement on FLSMD with 2% labels, but Section 5.2.2 states 4.76%. Fourth, encoder unfairness: Table 1 compares CTFS (DINOv2-S) against baselines using ResNet-101 (AEL, CPS variants), conflating architecture gains with backbone representation power.

“Notably, on the FLSMD dataset, our method achieves a 4.76% improvement even with only 2% of extremely scarce labeled data.”
paper · Section 5.2.2
“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.”
paper · Abstract
Evidence and comparison

The primary evidence (Table 1) shows CTFS (62.32% mIoU) outperforming UniMatch V2 (57.24%) on FLSMD with 2% labels. Since both use DINOv2-S encoders, this 5.08-point gap likely reflects genuine architectural improvements from collaborative teaching and MVRA. However, the paper lacks statistical significance testing (no variance across runs), and the new FSSG dataset (3,761 images, 11 categories) exhibits severe class imbalance (diver: 2% of samples) without specialized handling beyond the general reliability mechanism. Comparisons to SemiVL and Beyond-Pixels are less convincing given the encoder mismatch.

“CTFS (Ours) DINOv2-S 62.32 ... UniMatch V2 DINOv2-S 57.24”
paper · Table 1
“the steel frame category has the most samples (1229, 31%), while the diver category has the fewest (80, 2%), revealing the dataset's inherent long-tailed distribution.”
paper · Section 4
Reproducibility

Reproducibility is inadequate. No code repository URL is provided. While optimizer settings (AdamW, lr=5e-6 for encoder, 2e-4 for decoder) and grid size $m=32$ are specified, critical hyperparameters for the physics simulations—shadow intensity $\alpha$, attenuation factor $\gamma$, and warm-up epochs $E$—are not explicitly stated in the text. Training was performed on a single RTX 4090, but total training time, exact augmentation pipelines beyond the general description, and standard deviation across random seeds are omitted. The FSSG dataset is promised but no download link is visible in the provided text.

“We use the simple DPT as our semantic segmentation model... The learning rate for the pre-trained encoder is set to 5e-6, while that for the randomly initialized decoder is set to 2e-4.”
paper · Section 5.1
“experiments also show a 32×32 grid outperforms other sizes”
paper · Section 5.3.2
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 pseudo-labels, 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.

Challenge the Review

Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.

No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.