Research map / Label-efficient and robust learning / Knowledge distillation

CoDTS: Enhancing Sparsely Supervised Collaborative Perception with a Dual Teacher-Student Framework

AAAI 2025 · Label-efficient and robust learning › Knowledge distillation · approach: Few-shot and metric distillation · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CoDTS, a dual teacher-student framework for sparsely supervised collaborative 3D object detection that combines static-teacher high-quality pseudo labels with dynamic-teacher supplementary pseudo labels under a staged training strategy.

Problem

Collaborative 3D perception trained with sparse labels (one annotated object per agent per frame) suffers from a quality-quantity imbalance in pseudo labels produced by confidence-thresholded teacher-student frameworks.

Approach

Adaptive complementary learning: a static teacher with a high threshold (MFM) gives high-quality pseudo labels, while a dynamic EMA teacher with a per-batch K-Means-adapted threshold (SFM) recovers missed instances; neighbor anchor sampling (NAS) densifies pseudo labels, and a warm-up/refinement staged schedule trains student and dynamic teacher mutually.

Key results

• On DAIR-V2X with F-Cooper, CoDTS reaches 70.35 AP@0.5 and 50.53 AP@0.7 versus 54.26/35.39 for the static teacher baseline under sparse labels (ST row, AP@0.5 and AP@0.7 for the sparse ST baseline).
• CoDTS improves AP by over 20% at IoU 0.3 and 0.5 on DAIR-V2X and V2V4Real over sparse baselines, and by over 10% on OPV2V.
• On DAIR-V2X pseudo-label statistics with F-Cooper, CoDTS yields 12.33 pseudo labels per frame with MPR 0.49, versus 6.48 (MPR 0.66) for HINTED and 3.84 (MPR 0.75) for SSC3OD.

Measured effects reported in the paper

Surprising findings

• Applying dynamic threshold adaptation to the static teacher (MFM) made training unstable, attributed to the static teacher's weak sparse-label predictions.
• Directly stacking SFM after a separately trained first stage underperforms the end-to-end staged strategy, and training with only the refinement stage degrades accuracy notably (AP@0.7 42.16 vs 50.53 on F-Cooper).

Limitations

• Performance under sparse supervision still lags behind full supervision on several settings (e.g., DAIR-V2X at IoU 0.7).
• Dynamic threshold adaptation cannot be applied to the static teacher due to instability.
• Evaluated only with intermediate fusion detectors and mostly single-class (Car) datasets.

Experimental setup

Datasets: V2X-Sim OPV2V DAIR-V2X V2V4Real

Benchmarks: V2X-Sim OPV2V DAIR-V2X V2V4Real

Metrics: AP@0.3 AP@0.5 AP@0.7 false prediction ratio missing prediction ratio

Baselines: SSC3OD HINTED ST (static teacher) 3DIoUMatch HSSDA Full supervision

Builds on: HINTED SSC3OD OpenCOOD NMS K-Means

Closest papers by meaning

• Pre-Trained Vision-Language Models as Noisy Partial Annotators (AAAI 2025 · similarity 0.84)
• Collaborative Learning for Deep Neural Networks (NeurIPS 2018 · similarity 0.84)
• Mutual Learning for SAM Adaptation: A Dual Collaborative Network Framework for Source-Free Domain Transfer (ICML 2025 · similarity 0.84)
• What Makes Good Collaborative Views? Contrastive Mutual Information Maximization for Multi-Agent Perception (AAAI 2024 · similarity 0.84)
• NegoCollab: A Common Representation Negotiation Approach for Heterogeneous Collaborative Perception (NeurIPS 2025 · similarity 0.83)
• An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.83)