GradingNet: Towards Providing Reliable Supervisions for Weakly Supervised Object Detection by Grading the Box Candidates

Qifei Jia, Shikui Wei, Tao Ruan, Yufeng Zhao, Yao ZhaoAAAI 2021 · DOI

Abstract

Weakly Supervised Object Detection (WSOD) aims to train object detectors using only image-level labels. However, existing methods often suffer from "incomplete" pseudo-ground-truth boxes that focus on discriminative parts rather than the whole object. This paper introduces GradingNet, a two-stage framework that addresses this by grading candidate boxes based on their completeness and reliability. By incorporating a Boxes Grading Module (BGM) and an Informative Boosting Module (IBM), GradingNet achieves state-of-the-art performance on PASCAL VOC and MS COCO benchmarks.

Introduction

Weakly Supervised Object Detection (WSOD) is a critical task for scaling computer vision systems without the need for expensive instance-level annotations. Most current WSOD frameworks follow a two-stage pipeline: first, generating pseudo-ground-truth boxes via Multiple Instance Learning (MIL), and second, training a standard detector using these proposals. A persistent challenge is the "part-dominance" problem, where the model identifies only the most discriminative parts of an object (e.g., a bird's head) instead of the full extent. GradingNet proposes to move beyond binary labels by "grading" candidate boxes, allowing the second-stage detector to learn from varying levels of supervision quality.

Proposed Framework: GradingNet

The GradingNet architecture consists of two primary components designed to refine and utilize noisy candidate boxes.

Overview of the GradingNet framework, illustrating the transition from noisy WSOD outputs to graded labels and informative boosting.

1. Boxes Grading Module (BGM)

The BGM evaluates the quality of candidate boxes using two key principles:

  • Stability: By introducing pixel-level noise to images, the module identifies boxes that remain consistent across perturbations. Stable boxes are more likely to correspond to actual object features rather than background noise.
  • The Inclusion Principle: This principle leverages the hierarchical nature of objects. A "complete" box should ideally contain several smaller "part" boxes. Using an Intersection over Single (IoS) metric, the BGM constructs a graph where nodes represent stable boxes. A higher grade is assigned to boxes that encompass a greater number of constituent parts.
The Inclusion Principle illustrated through box candidates and their corresponding relationship graph, where nodes represent stable boxes and edges denote inclusion.

2. Informative Boosting Module (IBM)

Once boxes are graded, the IBM utilizes these "Graded-Labels" to train a more robust detector through three innovations:

  • Predicted Anchoring: Instead of standard grid-based anchors, the model predicts anchor locations based on the distribution of high-grade labels, focusing the search space on promising regions.
  • Grade-aware Loss: The regression loss is weighted by the reliability score (grade) of each box. This ensures that high-quality, complete boxes have a stronger influence on the model's parameters.
  • Area-balanced Sampling: To mitigate the imbalance between positive and negative samples, the module dynamically adjusts sampling strategies based on the spatial characteristics and quality of the proposals.

Experimental Results

GradingNet was evaluated on the PASCAL VOC 2007, 2012, and MS COCO 2014 datasets. It achieved state-of-the-art (SOTA) results across the board:

  • PASCAL VOC 2007: Reached 54.3% mAP, a significant improvement over previous leading methods like C-MIL (50.5%) and WSOD2 (53.6%).
  • PASCAL VOC 2012: Achieved 50.5% mAP.
  • MS COCO: Achieved 11.6% AP and 25.0% AP₅₀. While these numbers are lower than fully supervised models, they represent a new peak for the weakly supervised field on this difficult dataset.

Qualitative analysis confirms that GradingNet is more effective at capturing full object extents, reducing the common error of boxing only discriminative parts (e.g., just the head of a person or horse).

Conclusion

GradingNet represents a significant shift in WSOD by treating pseudo-labels as graded supervisions rather than absolute truths. By formalizing object completeness through the Inclusion Principle and stability through noise perturbation, the framework provides a more reliable foundation for training second-stage detectors. Future work will focus on improving performance in highly crowded scenes where overlapping objects complicate the hierarchical relationship between parts and wholes.

Summary updated

BibTeX
@inproceedings{jia2021gradingnet,
  author = {Qifei Jia and Shikui Wei and Tao Ruan and Yufeng Zhao and Yao Zhao},
  title = {{GradingNet: Towards Providing Reliable Supervisions for Weakly Supervised Object Detection by Grading the Box Candidates}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i2.16261},
}
Figure