Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Sparse Generation: Making Pseudo Labels Sparse for Point Weakly Supervised Object Detection on Low Data Volume

Chuyang Shang, Tian Ma, Wanzhu Ren, Yuancheng Li, Jiayi Yang

TL;DR

This work addresses the density and inefficiency of pseudo-label generation in point weakly supervised object detection under low data volume. It introduces Sparse Generation, a non-networked three-stage pipeline (Mapping, Mask, Regression) that converts dense pseudo labels into sparse representations and refines them with a regression step, aided by perspective-based matching (PADM) to recover missed instances. The approach relies on a lightweight mapping into 2D tensors, a mask-based aggregation to reduce localization bias, and a constrained regression trained with limited supervision, achieving substantial improvements over state-of-the-art on four datasets, including a newly introduced Bullet-Hole dataset for dense objects. Overall, Sparse Generation delivers higher-quality, sparser pseudo labels with lower computational cost, enabling effective one-epoch training in low-data regimes and offering practical gains for dense-object detection tasks.

Abstract

Existing pseudo label generation methods for point weakly supervised object detection are inadequate in low data volume and dense object detection tasks. We consider the generation of weakly supervised pseudo labels as the model's sparse output, and propose Sparse Generation as a solution to make pseudo labels sparse. The method employs three processing stages (Mapping, Mask, Regression), constructs dense tensors through the relationship between data and detector model, optimizes three of its parameters, and obtains a sparse tensor, thereby indirectly obtaining higher quality pseudo labels, and addresses the model's density problem on low data volume. Additionally, we propose perspective-based matching, which provides more rational pseudo boxes for prediction missed on instances. In comparison to the SOTA method, on four datasets (MS COCO-val, RSOD, SIMD, Bullet-Hole), the experimental results demonstrated a significant advantage.

Sparse Generation: Making Pseudo Labels Sparse for Point Weakly Supervised Object Detection on Low Data Volume

TL;DR

This work addresses the density and inefficiency of pseudo-label generation in point weakly supervised object detection under low data volume. It introduces Sparse Generation, a non-networked three-stage pipeline (Mapping, Mask, Regression) that converts dense pseudo labels into sparse representations and refines them with a regression step, aided by perspective-based matching (PADM) to recover missed instances. The approach relies on a lightweight mapping into 2D tensors, a mask-based aggregation to reduce localization bias, and a constrained regression trained with limited supervision, achieving substantial improvements over state-of-the-art on four datasets, including a newly introduced Bullet-Hole dataset for dense objects. Overall, Sparse Generation delivers higher-quality, sparser pseudo labels with lower computational cost, enabling effective one-epoch training in low-data regimes and offering practical gains for dense-object detection tasks.

Abstract

Existing pseudo label generation methods for point weakly supervised object detection are inadequate in low data volume and dense object detection tasks. We consider the generation of weakly supervised pseudo labels as the model's sparse output, and propose Sparse Generation as a solution to make pseudo labels sparse. The method employs three processing stages (Mapping, Mask, Regression), constructs dense tensors through the relationship between data and detector model, optimizes three of its parameters, and obtains a sparse tensor, thereby indirectly obtaining higher quality pseudo labels, and addresses the model's density problem on low data volume. Additionally, we propose perspective-based matching, which provides more rational pseudo boxes for prediction missed on instances. In comparison to the SOTA method, on four datasets (MS COCO-val, RSOD, SIMD, Bullet-Hole), the experimental results demonstrated a significant advantage.
Paper Structure (10 sections, 10 equations, 2 figures, 2 tables)

This paper contains 10 sections, 10 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Sparse Generation
  3. Mapping stage
  4. Mask stage
  5. Regression Stage
  6. Perspective average distance matching
  7. Experiments
  8. Dataset and experiment detail
  9. Experiment Results on MS COCO, Bullet-Hole, RSOD and SIMD Datasets
  10. Conclusion

Figures (2)

  • Figure 1: The pipeline of Sparse Generation. It uses non-networked approach and direct regression on pseudo labels.
  • Figure 2: Experiment result on MS COCO-val lin2014microsoft, SIMD haroon2020multisized, RSOD li2020object and Bullet-Hole datasets.