Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Co-Learning: Towards Semi-Supervised Object Detection with Road-side Cameras

Jicheng Yuan, Anh Le-Tuan, Ali Ganbarov, Manfred Hauswirth, Danh Le-Phuoc

TL;DR

The paper tackles the challenge of semi-supervised object detection for road-side camera data under data-scarce conditions. It introduces Co-Learning, a teacher–student framework that uses data curation, annotation-alignment, dynamic pseudo-labels, and multi-head student configurations to mitigate pseudo-label noise and task misalignment. On the AI City Challenge Track 2 dataset, starting from only $10\%$ labeled data, the approach achieves robust $mAP$ gains, reaching $36.5$ with unlabeled data and annotation alignment, surpassing an oracle model relying solely on pseudo labels by $0.4\%$. The work demonstrates that label-consistent SSOD can significantly reduce annotation costs while maintaining strong detection performance, with clear paths to edge-device deployment and integration with visual-language cues for enhanced practicality.

Abstract

Recently, deep learning has experienced rapid expansion, contributing significantly to the progress of supervised learning methodologies. However, acquiring labeled data in real-world settings can be costly, labor-intensive, and sometimes scarce. This challenge inhibits the extensive use of neural networks for practical tasks due to the impractical nature of labeling vast datasets for every individual application. To tackle this, semi-supervised learning (SSL) offers a promising solution by using both labeled and unlabeled data to train object detectors, potentially enhancing detection efficacy and reducing annotation costs. Nevertheless, SSL faces several challenges, including pseudo-target inconsistencies, disharmony between classification and regression tasks, and efficient use of abundant unlabeled data, especially on edge devices, such as roadside cameras. Thus, we developed a teacher-student-based SSL framework, Co-Learning, which employs mutual learning and annotation-alignment strategies to adeptly navigate these complexities and achieves comparable performance as fully-supervised solutions using 10\% labeled data.

Co-Learning: Towards Semi-Supervised Object Detection with Road-side Cameras

TL;DR

The paper tackles the challenge of semi-supervised object detection for road-side camera data under data-scarce conditions. It introduces Co-Learning, a teacher–student framework that uses data curation, annotation-alignment, dynamic pseudo-labels, and multi-head student configurations to mitigate pseudo-label noise and task misalignment. On the AI City Challenge Track 2 dataset, starting from only labeled data, the approach achieves robust gains, reaching with unlabeled data and annotation alignment, surpassing an oracle model relying solely on pseudo labels by . The work demonstrates that label-consistent SSOD can significantly reduce annotation costs while maintaining strong detection performance, with clear paths to edge-device deployment and integration with visual-language cues for enhanced practicality.

Abstract

Recently, deep learning has experienced rapid expansion, contributing significantly to the progress of supervised learning methodologies. However, acquiring labeled data in real-world settings can be costly, labor-intensive, and sometimes scarce. This challenge inhibits the extensive use of neural networks for practical tasks due to the impractical nature of labeling vast datasets for every individual application. To tackle this, semi-supervised learning (SSL) offers a promising solution by using both labeled and unlabeled data to train object detectors, potentially enhancing detection efficacy and reducing annotation costs. Nevertheless, SSL faces several challenges, including pseudo-target inconsistencies, disharmony between classification and regression tasks, and efficient use of abundant unlabeled data, especially on edge devices, such as roadside cameras. Thus, we developed a teacher-student-based SSL framework, Co-Learning, which employs mutual learning and annotation-alignment strategies to adeptly navigate these complexities and achieves comparable performance as fully-supervised solutions using 10\% labeled data.

Paper Structure

This paper contains 11 sections, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Object Detection
  4. Semi-supverised Object Detection
  5. Methodology and Experiments
  6. Data Curation and Annotation Alignment
  7. Overall Training Platform
  8. Experimental setup and Results
  9. Further Analysis
  10. Conclusions
  11. Acknowledgements

Figures (2)

  • Figure 1: Overview of the proposed Co-Learning framework: three modules to address the inconsistency in SSOD on computing-power-constrained devices, $\left(i\right)$ Pseudo-labels are determined by dynamic thresholds; $\left(ii\right)$ Consistent pseudo-labels across teacher and student networks contribute to aligning and calibrating regression quality; $\left(iii\right)$ Multi-head student networks reassign anchors based on matching cost.
  • Figure 2: Qualititive Analysis of the proposed Co-Learning solution on the 7th AI CITY CHALLENGE Track 2 dataset. The base detector is Faster-RCNN with ResNet-50 as the feature extractor for all experiments to ensure a fair comparison. (DETs: Detected Results, GTs: Ground Truths)