Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning to Borrow Features for Improved Detection of Small Objects in Single-Shot Detectors

Richard Schmit

TL;DR

The paper tackles the persistent difficulty of detecting small objects in single-shot detectors by enabling shallow feature maps to borrow discriminative information from deeper, more semantic layers within the same class. It introduces a cross-layer borrowing pipeline consisting of three blocks: Feature Matching Block (FMB) to identify semantically similar descriptors, Feature Representing Block (FRB) to assemble borrowed features, and Feature Fusion Block (FFB) to integrate borrowed features with original and contextual information, all built on the SSD framework to maintain real-time performance. The approach yields improved small-object detection accuracy over baselines and demonstrates the potential of cross-layer feature borrowing for robust detection in complex scenes. The work also discusses anchor design refinements and broader conceptual connections to multi-source priors across domains, suggesting broad applicability and directions for future extensions.

Abstract

Detecting small objects remains a significant challenge in single-shot object detectors due to the inherent trade-off between spatial resolution and semantic richness in convolutional feature maps. To address this issue, we propose a novel framework that enables small object representations to "borrow" discriminative features from larger, semantically richer instances within the same class. Our architecture introduces three key components: the Feature Matching Block (FMB) to identify semantically similar descriptors across layers, the Feature Representing Block (FRB) to generate enhanced shallow features through weighted aggregation, and the Feature Fusion Block (FFB) to refine feature maps by integrating original, borrowed, and context information. Built upon the SSD framework, our method improves the descriptive capacity of shallow layers while maintaining real-time detection performance. Experimental results demonstrate that our approach significantly boosts small object detection accuracy over baseline methods, offering a promising direction for robust object detection in complex visual environments.

Learning to Borrow Features for Improved Detection of Small Objects in Single-Shot Detectors

TL;DR

The paper tackles the persistent difficulty of detecting small objects in single-shot detectors by enabling shallow feature maps to borrow discriminative information from deeper, more semantic layers within the same class. It introduces a cross-layer borrowing pipeline consisting of three blocks: Feature Matching Block (FMB) to identify semantically similar descriptors, Feature Representing Block (FRB) to assemble borrowed features, and Feature Fusion Block (FFB) to integrate borrowed features with original and contextual information, all built on the SSD framework to maintain real-time performance. The approach yields improved small-object detection accuracy over baselines and demonstrates the potential of cross-layer feature borrowing for robust detection in complex scenes. The work also discusses anchor design refinements and broader conceptual connections to multi-source priors across domains, suggesting broad applicability and directions for future extensions.

Abstract

Detecting small objects remains a significant challenge in single-shot object detectors due to the inherent trade-off between spatial resolution and semantic richness in convolutional feature maps. To address this issue, we propose a novel framework that enables small object representations to "borrow" discriminative features from larger, semantically richer instances within the same class. Our architecture introduces three key components: the Feature Matching Block (FMB) to identify semantically similar descriptors across layers, the Feature Representing Block (FRB) to generate enhanced shallow features through weighted aggregation, and the Feature Fusion Block (FFB) to refine feature maps by integrating original, borrowed, and context information. Built upon the SSD framework, our method improves the descriptive capacity of shallow layers while maintaining real-time detection performance. Experimental results demonstrate that our approach significantly boosts small object detection accuracy over baseline methods, offering a promising direction for robust object detection in complex visual environments.
Paper Structure (10 sections, 1 equation, 3 figures, 1 table)

This paper contains 10 sections, 1 equation, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Model
  4. Feature Matching Block
  5. Feature Representing Block
  6. Feature Fusion Block
  7. Experiment
  8. Anchor Design
  9. Discussion
  10. Conclusion

Figures (3)

  • Figure 1: Examples from COCO dataset showing multiple instances with various sizes of the same class.
  • Figure 2: The overall architecture of the proposed architecture, where only the layers aimed for detection are displayed for better visualization.
  • Figure 3: The diagrams of feature matching block (FMB), feature representing block (FRB) and feature fusion block (FFB). Here, "$Concat$" denotes the concatenation along the first dimension and "$Concat\_c$" denotes the concatenation along the channel dimension.