Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-Grid Redundant Bounding Box Annotation for Accurate Object Detection

Solomon Negussie Tesema, El-Bay Bourennane

TL;DR

This work introduces MultiGridDet, a lightweight, one-stage object detector that extends grid-based annotation by assigning multiple grids to each object to encourage multi-perspective predictions and tighter bounding boxes. It integrates a DenseYOLO-inspired output head and a four-part loss (class, anchor, coordinate, and objectness) to train with a novel offline copy-paste augmentation that drastically expands the training set. Empirical results show improved VOC performance over prior YOLO variants and competitive COCO results, particularly for large objects, while achieving faster inference than YOLOv3. The approach aims to address small-object limitations and imbalance across grid cells, with future directions toward even finer-scale outputs and broader tasks like tracking and segmentation.

Abstract

Modern leading object detectors are either two-stage or one-stage networks repurposed from a deep CNN-based backbone classifier network. YOLOv3 is one such very-well known state-of-the-art one-shot detector that takes in an input image and divides it into an equal-sized grid matrix. The grid cell having the center of an object is the one responsible for detecting the particular object. This paper presents a new mathematical approach that assigns multiple grids per object for accurately tight-fit bounding box prediction. We also propose an effective offline copy-paste data augmentation for object detection. Our proposed method significantly outperforms some current state-of-the-art object detectors with a prospect for further better performance.

Multi-Grid Redundant Bounding Box Annotation for Accurate Object Detection

TL;DR

This work introduces MultiGridDet, a lightweight, one-stage object detector that extends grid-based annotation by assigning multiple grids to each object to encourage multi-perspective predictions and tighter bounding boxes. It integrates a DenseYOLO-inspired output head and a four-part loss (class, anchor, coordinate, and objectness) to train with a novel offline copy-paste augmentation that drastically expands the training set. Empirical results show improved VOC performance over prior YOLO variants and competitive COCO results, particularly for large objects, while achieving faster inference than YOLOv3. The approach aims to address small-object limitations and imbalance across grid cells, with future directions toward even finer-scale outputs and broader tasks like tracking and segmentation.

Abstract

Modern leading object detectors are either two-stage or one-stage networks repurposed from a deep CNN-based backbone classifier network. YOLOv3 is one such very-well known state-of-the-art one-shot detector that takes in an input image and divides it into an equal-sized grid matrix. The grid cell having the center of an object is the one responsible for detecting the particular object. This paper presents a new mathematical approach that assigns multiple grids per object for accurately tight-fit bounding box prediction. We also propose an effective offline copy-paste data augmentation for object detection. Our proposed method significantly outperforms some current state-of-the-art object detectors with a prospect for further better performance.
Paper Structure (13 sections, 20 equations, 5 figures, 3 tables)

This paper contains 13 sections, 20 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Multi-Grid Assignment
  4. Training
  5. The Detection Network: MultiGridDet
  6. The Loss function
  7. Class prediction loss (error)
  8. Anchor prediction loss (error)
  9. Coordinate prediction loss (error)
  10. Objectness confidence loss
  11. Data Augmentation
  12. Experiment
  13. Conclusion

Figures (5)

  • Figure 1: Multi-grid assignment
  • Figure 2: Ground-truth encoding
  • Figure 3: Coordinate activation function plot with different $\beta$ values
  • Figure 4: Sample Offline Copy-Paste Generated Artificial Images
  • Figure 5: Sample MultiGridDet output on randomly selected Pascal VOC 2007 test set images. As seen from the figure the first row shows six input images, whereas the second row shows the prediction of the network before non-max-suppression (NMS) and the last row shows the final bounding box prediction of MultiGridDet on the input image after NMS thresholding.