Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Practical Insights into Semi-Supervised Object Detection Approaches

Chaoxin Wang, Bharaneeshwar Balasubramaniyam, Anurag Sangem, Nicolais Guevara, Doina Caragea

TL;DR

This work benchmarks three state-of-the-art SSOD approaches—MixPL, Semi-DETR, and Consistent-Teacher—under a fixed per-class labeling regime defined by $k$-shot, across MS-COCO, Pascal VOC, and a Beetle dataset. By standardizing the amount of labeled data per category and reporting both detection accuracy ($mAP$) and deployment metrics (inference latency and model footprint), the study reveals how performance scales with $k$ and how architecture choices trade off accuracy against computation and memory requirements. The results show that transformer-based methods (MixPL, Semi-DETR) achieve higher peak accuracy, especially in mid-to-high data regimes, but incur larger latency and larger footprints, while Consistent-Teacher is faster and lighter and can be competitive in very low-data scenarios. The findings provide practical guidance on selecting SSOD methods based on dataset complexity and deployment constraints, and call for broader reporting of resource metrics in SSOD studies.

Abstract

Learning in data-scarce settings has recently gained significant attention in the research community. Semi-supervised object detection(SSOD) aims to improve detection performance by leveraging a large number of unlabeled images alongside a limited number of labeled images(a.k.a.,few-shot learning). In this paper, we present a comprehensive comparison of three state-of-the-art SSOD approaches, including MixPL, Semi-DETR and Consistent-Teacher, with the goal of understanding how performance varies with the number of labeled images. We conduct experiments using the MS-COCO and Pascal VOC datasets, two popular object detection benchmarks which allow for standardized evaluation. In addition, we evaluate the SSOD approaches on a custom Beetle dataset which enables us to gain insights into their performance on specialized datasets with a smaller number of object categories. Our findings highlight the trade-offs between accuracy, model size, and latency, providing insights into which methods are best suited for low-data regimes.

Practical Insights into Semi-Supervised Object Detection Approaches

TL;DR

This work benchmarks three state-of-the-art SSOD approaches—MixPL, Semi-DETR, and Consistent-Teacher—under a fixed per-class labeling regime defined by -shot, across MS-COCO, Pascal VOC, and a Beetle dataset. By standardizing the amount of labeled data per category and reporting both detection accuracy () and deployment metrics (inference latency and model footprint), the study reveals how performance scales with and how architecture choices trade off accuracy against computation and memory requirements. The results show that transformer-based methods (MixPL, Semi-DETR) achieve higher peak accuracy, especially in mid-to-high data regimes, but incur larger latency and larger footprints, while Consistent-Teacher is faster and lighter and can be competitive in very low-data scenarios. The findings provide practical guidance on selecting SSOD methods based on dataset complexity and deployment constraints, and call for broader reporting of resource metrics in SSOD studies.

Abstract

Learning in data-scarce settings has recently gained significant attention in the research community. Semi-supervised object detection(SSOD) aims to improve detection performance by leveraging a large number of unlabeled images alongside a limited number of labeled images(a.k.a.,few-shot learning). In this paper, we present a comprehensive comparison of three state-of-the-art SSOD approaches, including MixPL, Semi-DETR and Consistent-Teacher, with the goal of understanding how performance varies with the number of labeled images. We conduct experiments using the MS-COCO and Pascal VOC datasets, two popular object detection benchmarks which allow for standardized evaluation. In addition, we evaluate the SSOD approaches on a custom Beetle dataset which enables us to gain insights into their performance on specialized datasets with a smaller number of object categories. Our findings highlight the trade-offs between accuracy, model size, and latency, providing insights into which methods are best suited for low-data regimes.
Paper Structure (24 sections, 1 figure, 5 tables)

This paper contains 24 sections, 1 figure, 5 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Semi-Supervised Object Detection
  4. Few-Shot Object Detection
  5. Semi-Supervised Few-Shot Object Detection
  6. Problem Statement
  7. Methodology
  8. Few-Shot Sampling Strategy
  9. SSOD Approaches
  10. Experimental Setup
  11. Model Training
  12. Datasets
  13. Evaluation Metrics
  14. Results & Discussion
  15. Performance Analysis Across Datasets
  16. ...and 9 more sections

Figures (1)

  • Figure 1: Performance comparisons with the number of $k$-shots across MixPL, Semi-DETR, and Consistent-Teacher on MS-COCO, Pascal VOC and Beetle datasets.