Collaborative Teacher-Student Network Models

Updated 8 July 2026

The paper introduces a two-stage framework that leverages simulation teaching and practical learning to transfer occlusion-robust features from full-body to occluded domains.
CoSD-TSNet defines collaborative knowledge transfer via shared branches, multi-task coupling, and pseudo-label generation in both re-identification and detection settings.
Empirical results demonstrate significant performance gains on occluded re-identification benchmarks and effective cross-modal detection using single-modality annotations.

Collaborative Teacher-Student Network (CoSD-TSNet) denotes a class of teacher-student architectures in which knowledge transfer is explicitly collaborative, but the name is not used in a single uniform way across the cited literature. In the provided sources, the designation is applied most concretely to two different systems: a two-stage framework for occluded person re-identification that performs “simulation teaching” on full-body person data and “practical learning” on real occluded data (Zhuo et al., 2019), and a Single- and Dual-Modality Collaborative Teacher-Student Network used inside DOD-SA for infrared-visible decoupled object detection with single-modality annotations (Jin et al., 14 Aug 2025). Related frameworks on collaborative online distillation and uncertainty-aware teacher-student learning are methodologically adjacent, but they are not literally the same named method (Sun et al., 2021, Si et al., 2023).

1. Scope and terminological usage

The term CoSD-TSNet does not denote a single canonical architecture across all of the papers on arXiv. In the occluded person re-identification setting, it refers to a two-stage teacher-student framework designed to cope with two core difficulties: the lack of large-scale occluded training data and the fact that occlusions corrupt identity features (Zhuo et al., 2019). In DOD-SA, the same label is used for a collaborative single- and dual-modality teacher-student network that learns decoupled detections in both infrared and visible modalities while only one modality is annotated (Jin et al., 14 Aug 2025).

Source	Domain	Usage of CoSD-TSNet
(Zhuo et al., 2019)	Occluded person re-identification	Two-stage teacher-student framework with “simulation teaching” and “practical learning”
(Jin et al., 14 Aug 2025)	Infrared-visible object detection	Single- and Dual-Modality Collaborative Teacher-Student Network with SM-Branch and DMD-Branch
(Sun et al., 2021, Si et al., 2023)	Related distillation / noisy-label learning	Conceptually close, but not literally the same named method

This terminological variation matters. A common misconception is to treat CoSD-TSNet as a single fixed design. The cited material instead shows a recurrent pattern: a teacher produces transferable supervision under a difficult supervision regime, while collaboration is introduced through shared branches, multi-task coupling, peer exchange, or cross-modality guidance. A plausible implication is that CoSD-TSNet is best understood as a family resemblance term rather than a uniquely standardized model name.

2. Occluded person re-identification: staged transfer from full-body to occluded domains

In occluded person re-identification, the framework is built around a staged transfer process from abundant full-body person data to scarce real occluded data (Zhuo et al., 2019). The key idea is to first learn an occlusion-robust representation in a teacher stage and then transfer that representation to a student stage trained on real occluded person data. The paper states that the teacher stage performs “simulation teaching,” while the student stage performs “practical learning.”

The teacher is not treated as a standard full-body re-identification model. It is explicitly taught to behave like an occluded-person model by means of a co-saliency network and a cross-domain simulator. The student then inherits this learned basis and adapts it to real occluded images. Knowledge transfer occurs in two ways. First, the student inherits the teacher’s learned parameters or learned occlusion-robust representation as the starting point. Second, the teacher’s co-saliency branch is reused to generate saliency masks for the real occluded person data, replacing the weaker masks from the initial salient object detector. This means that the teacher transfers not only identity-discriminative features but also improved pseudo-saliency annotations.

The rationale of the design is tightly coupled to the data regime. The framework assumes access to a large full-body person dataset for teacher training; in the paper this is MARS. It also assumes that the co-saliency masks used in teacher training are pseudo-labels from an existing salient object detector rather than manual annotations. The approach therefore addresses scarce occluded supervision by synthesizing difficult cases during teacher training and then using the teacher to bootstrap the student.

3. Re-identification architecture, simulator, and optimization

At the architectural level, the teacher network is a co-saliency network with one shared backbone and two collaborative heads (Zhuo et al., 2019). The backbone is a ResNet-50 feature extractor divided into five convolutional blocks. On top of this backbone, one branch is a classification branch for identity recognition and the other is a co-saliency branch for guiding the network to highlight meaningful parts without any manual annotation. The two branches are collaborative because they share the same backbone and optimize it toward a common objective: emphasize meaningful human body parts while retaining identity-discriminative power.

The co-saliency branch is implemented as four CS blocks. Each CS block upsamples intermediate features using a deconvolution layer with bilinear interpolation, aligns channels using parallel $1\times1$ convolutions, fuses features by pixel-wise summation, and then refines them with two $3\times3$ convolutions. The classification branch feeds semantic identity information into the shared representation, while the co-saliency branch encodes spatial location cues for body parts.

The cross-domain simulator is the mechanism that makes the teacher stage mimic the occluded domain even though it is trained on full-body person data. During data loading, it randomly selects a subset of full-body images and generates artificial occlusions by pasting a background patch onto the image. The same patch is also pasted into the corresponding saliency mask as a black region. This produces a simulated occluded image $I_s$ , a modified saliency mask $S_s$ , and an occlusion/non-occlusion binary label. Crucially, the probability of applying this occlusion grows with training iterations: it is initialized to $p\leftarrow 0$ , and after each epoch it is updated according to the epoch-to- $\text{epoch}_{max}$ schedule. The paper states that this gradual schedule reduces the domain gap between full-body and occluded data and lets the teacher progressively learn occlusion-robust features.

The loss design is multi-task. Let $f(\bullet)$ be the shared backbone, $g(\bullet)$ the identity classifier, $h(\bullet)$ the co-saliency predictor, and $b(\bullet)$ the occluded/non-occluded binary classifier. The co-saliency-network loss is

$3\times3$ 0

where $3\times3$ 1 is the classification loss and $3\times3$ 2 is the co-saliency loss, with $3\times3$ 3 because identity classification is the main re-identification objective. The simulator adds an occlusion/non-occlusion binary classification loss, and the classification branch becomes a multi-task branch:

$3\times3$ 4

with $3\times3$ 5. The overall teacher-stage objective across the two domains is

$3\times3$ 6

where $3\times3$ 7 because the framework considers the full-body domain and the occluded domain or simulated domain. The paper is explicit that the “domain transfer” effect is not a separate explicit adversarial loss; instead, it comes from the simulator’s progressive occlusion generation plus the OBC auxiliary task.

4. Re-identification benchmarks, results, and ablations

The experimental setup uses Occluded-REID, Partial-REID, P-DukeMTMC-reID, and P-ETHZ as occluded-person benchmarks, with MARS as the large-scale full-body training source for the teacher (Zhuo et al., 2019). The evaluation protocol is the standard re-identification one: queries are occluded person images and galleries are full-body person images. The paper reports CMC rank-1, rank-2, rank-5, and mAP, and for saliency quality it reports Precision, Recall, and F-measure with $3\times3$ 8. The implementation uses PyTorch with ResNet-50 as backbone, Adam optimization, learning rate $3\times3$ 9 for the backbone and $I_s$ 0 for the branches, $I_s$ 1, batch size 8, and 50K iterations. Input images are resized to $I_s$ 2 and randomly cropped to $I_s$ 3.

Benchmark	Reported “Ours” result	Metric subset shown
Occluded-REID	73.67% rank-1, 92.87% rank-5	rank-1, rank-5
Partial-REID	82.67% rank-1, 97.00% rank-5	rank-1, rank-5
P-DukeMTMC-reID	51.42% rank-1, 69.72% rank-5	rank-1, rank-5
P-ETHZ	62.86% rank-1, 85.24% rank-5	rank-1, rank-5

In the unsupervised setting on Partial-REID, the method achieves 69.2% rank-1, 85.8% rank-5, and 93.3% rank-10, outperforming SCPNet-a at 68.3% rank-1 and AFPB at 51.7% rank-1. Across the broader comparison, the method is generally best or near-best against classical re-identification methods such as XQDA, GOG, NullSpace, DGD, SVDNet, and REDA, and also against deep baselines such as PCB, MLFN, ResNet-mid, and AFPB. The authors note that P-DukeMTMC-reID is larger-scale and the method is less dominant there, suggesting that the approach is particularly effective when occluded labeled data are scarce.

The ablation study isolates the contribution of each component. Removing the student stage leads to much worse performance, showing that adaptation to real occluded data matters. Adding the co-saliency branch to the classifier improves over classification alone, confirming that saliency guidance helps the backbone focus on body parts rather than occlusions. Adding the cross-domain simulator yields large additional gains: the paper reports rank-1 improvements of 6.53%, 8.00%, 3.32%, and 13.5% on the four datasets in the teacher-only setting, with similar gains in the supervised setting. Adding the OBC loss then brings a smaller but consistent boost over the simulator without OBC. Separate analysis of saliency prediction shows that the co-saliency branch outperforms DSS, NLDF, and PiCA on occluded-person saliency masks, especially in recall and F-score. The simulator is also ablated by varying the occlusion probability among 0, 1, and the proposed growing schedule, and the growing schedule is reported as best.

Several conceptual assumptions delimit the method. Artificial occlusions are generated by pasting background patches, so the simulated occlusions are synthetic and may not fully capture all real-world occlusion patterns. In the supervised case, the student depends on the availability of some real occluded data for adaptation; in the unsupervised case, the student stage is skipped and performance is lower.

5. Infrared-visible detection: single- and dual-modality CoSD-TSNet

In DOD-SA, CoSD-TSNet addresses a different problem: the model must output decoupled detections in both infrared and RGB, but training annotations are available for only one modality (Jin et al., 14 Aug 2025). The central challenge is that IR and visible images are often misaligned, annotating both modalities is expensive, and decoupled detection still requires both outputs at inference. The framework addresses this by using one labeled modality to train directly, generating pseudo-labels for the unlabeled modality, transferring knowledge from labeled to unlabeled modality, and training a branch that outputs two separate modality-specific boxes for each object.

The architecture consists of a single-modality branch and a dual-modality decoupled branch. The SM-Branch processes one modality at a time,

$I_s$ 4

and outputs one classification result $I_s$ 5 and one modality-specific box $I_s$ 6. The DMD-Branch simultaneously consumes both IR and RGB,

$I_s$ 7

and outputs one classification result $I_s$ 8 together with decoupled position results for both modalities,

$I_s$ 9

Both teacher and student models include IR Stream and RGB Stream with unshared parameters, an SP-Decoder for the SM-Branch, and a DP-Decoder for the DMD-Branch, with the latter adding a feature fusion module before decoding.

Teacher-student supervision is explicit. The teacher is updated by EMA and produces pseudo-labels for the unlabeled modality; the student is trained with supervision from both true labels and pseudo-labels. For the labeled modality, the student’s SM-Branch takes labeled IR input $S_s$ 0 and predicts $S_s$ 1 with supervision from $S_s$ 2 using

$S_s$ 3

For the unlabeled RGB modality, the student sees strong augmentation $S_s$ 4, the teacher sees weak augmentation $S_s$ 5, and the teacher outputs RGB pseudo-labels $S_s$ 6. These supervise the student via

$S_s$ 7

Cross-modality knowledge transfer occurs in two complementary ways. The first is teacher-student transfer from the labeled modality to the unlabeled modality through pseudo-label generation. The second is branch collaboration: the SM-Branch is used to guide the training of the DMD-Branch, so knowledge first learned in single-modality detection is reused to supervise the harder decoupled two-modality task. In Stage 2, the DMD-Branch is supervised with

$S_s$ 8

where the paired supervision is formed by matching IR ground truth with RGB pseudo-labels.

6. PaST, PLA, and empirical behavior in DOD-SA

DOD-SA introduces a Progressive and Self-Tuning Training Strategy (PaST) that trains the model in three stages (Jin et al., 14 Aug 2025). Stage 1 pretrains the SM-Branch. During burn-in, the student SM-Branch is trained on labeled IR data for $S_s$ 9 epochs. In the subsequent teacher-student mutual learning phase, the teacher generates RGB pseudo-labels for unlabeled data and the student is trained with $p\leftarrow 0$ 0, where $p\leftarrow 0$ 1 is linearly increased from 0 to 1.0. Stage 2 retains the SM-Branch and additionally activates the DMD-Branch in the student; the IR Stream parameters are copied into the RGB Stream, and the total stage loss is $p\leftarrow 0$ 2, running for $p\leftarrow 0$ 3 epochs. Stage 3 activates only the DMD-Branch in both teacher and student, uses the same paired loss for $p\leftarrow 0$ 4 epochs, and further refines the DMD-Branch. After training, only the teacher network’s DMD-Branch structure is retained.

PLA is the module that aligns and pairs labels across modalities under misalignment. It has three parts: Pseudo Label Filter (PLF), Shape-aware Dual-modality Label Matcher (SDLM), and Dynamic Label Correction (DLC). In PLF, each RGB pseudo-label receives confidence

$p\leftarrow 0$ 5

and the threshold is batch-adaptive,

$p\leftarrow 0$ 6

with $p\leftarrow 0$ 7 and $p\leftarrow 0$ 8 computed from the batch scores. The paper states that this is better than a fixed threshold because pseudo-label quality varies by batch.

SDLM pairs IR ground-truth boxes with RGB pseudo-labels using a shape-aware search region rather than naive one-to-one matching. Each box is

$p\leftarrow 0$ 9

and the search region around IR box $\text{epoch}_{max}$ 0 is defined using a rotated region controlled by hyperparameter $\text{epoch}_{max}$ 1. Within this region, the visible candidate with maximum IoU is selected. DLC then maintains a dynamically updated bag of label pairs. At the first epoch of Stage 2, unmatched IR boxes are copied into the visible modality and combined with matched pairs; at later epochs, each pair is updated if a better RGB pseudo-label is found. The effect is that PLA not only filters and matches labels but also dynamically corrects them over training.

The empirical evaluation is conducted on DroneVehicle, with 17,990 training pairs and 1,469 test pairs, using [email protected]. When trained with IR labels only, DOD-SA achieves 80.41% IR-test mAP and 78.87% RGB-test mAP. When trained with RGB labels only, it achieves 77.96% IR-test mAP and 78.19% RGB-test mAP. The paper states that this outperforms single-modality detectors trained under the same condition and also dual-modality methods requiring full IR+RGB labels; it also reports that DOD-SA surpasses DPDETR, a fully supervised decoupled method, despite using only single-modality labels. The ablations support each core design choice: each stage of PaST improves performance, Stage 3 gives an additional boost through self-refinement, adaptive thresholding in PLF outperforms fixed thresholds, SDLM is better than naive IoU matching and much better than no pseudo labels, and DLC outperforms simple label correction or no dynamic update.

Two additional papers clarify the conceptual neighborhood of CoSD-TSNet while also showing that the name is not standardized. CTSL-MKT proposes “Collaborative Teacher-Student Learning via Multiple Knowledge Transfer,” a unified knowledge-distillation framework in which two peer networks act as both teacher and student for each other while also self-learning from their own earlier predictions (Sun et al., 2021). Its training has a pre-training stage and a collaborative stage, and each peer is optimized with

$\text{epoch}_{max}$ 2

The transferred knowledge includes response-based transfer via KL divergence and relation-based transfer through distance-wise and angle-wise similarity. The paper is explicit that CTSL-MKT is not literally the same named method as CoSD-TSNet, but is “very closely related conceptually,” and can be viewed as a precursor-style framework or close cousin of a collaborative self-distillation teacher-student network.

A second neighboring line appears in distantly supervised named entity recognition. The paper “Improving the Robustness of Distantly-Supervised Named Entity Recognition via Uncertainty-Aware Teacher Learning and Student-Student Collaborative Learning” is summarized as CoSD-TSNet / CENSOR in the provided material, but the actual paper title uses CENSOR rather than CoSD-TSNet (Si et al., 2023). Its architecture uses two teacher-student pairs with RoBERTa-base and DistilRoBERTa-base, Uncertainty-Aware Teacher Learning to select labels that are both high-confidence and low-uncertainty, Student-Student Collaborative Learning to exchange small-loss pseudo labels, and EMA teacher updates

$\text{epoch}_{max}$ 3

This again aligns with the broader collaborative teacher-student pattern, but the nomenclature differs.

These distinctions delimit what CoSD-TSNet should and should not be taken to mean. It should not be assumed to imply only classical offline knowledge distillation, because the cited systems include fixed teacher-to-student transfer, EMA-updated pseudo-labeling, SM-to-DMD branch supervision, and peer-style collaboration in related frameworks. It also should not be assumed to imply a specific adversarial domain-transfer mechanism; in the occluded re-identification case, the transfer effect is explicitly attributed to progressive synthetic occlusion and the OBC auxiliary task rather than an adversarial loss. A plausible implication is that the most stable meaning of CoSD-TSNet is architectural rather than nominal: a collaborative teacher-student arrangement designed to make weak, partial, or noisy supervision usable in a harder target setting.