Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Object Detection Performance for Small Objects through Synthetic Data Generation and Proportional Class-Balancing Technique: A Comparative Study in Industrial Scenarios

Jibinraj Antony, Vinit Hegiste, Ali Nazeri, Hooman Tavakoli, Snehal Walunj, Christiane Plociennik, Martin Ruskowski

TL;DR

The paper addresses the poor performance of object detectors on small industrial objects due to data scarcity and imbalance. It introduces synthetic data generation using CAD models and a proportional class-balancing scheme to balance small-object samples and improve anchor matching, evaluated on YOLOv5, YOLOv7, and SSD in an industrial assembly scenario. Across datasets, a balanced mix where synthetic data equals about half of real data (DS-3) yields the best small-object AP gains, with YOLOv5 delivering the strongest overall results; excessive synthetic data can hurt performance due to non-photo-realistic rendering. The work highlights practical trade-offs between realism and data diversity, suggesting that improved synthetic realism could further boost small-object detection in industry.

Abstract

Object Detection (OD) has proven to be a significant computer vision method in extracting localized class information and has multiple applications in the industry. Although many of the state-of-the-art (SOTA) OD models perform well on medium and large sized objects, they seem to under perform on small objects. In most of the industrial use cases, it is difficult to collect and annotate data for small objects, as it is time-consuming and prone to human errors. Additionally, those datasets are likely to be unbalanced and often result in an inefficient model convergence. To tackle this challenge, this study presents a novel approach that injects additional data points to improve the performance of the OD models. Using synthetic data generation, the difficulties in data collection and annotations for small object data points can be minimized and to create a dataset with balanced distribution. This paper discusses the effects of a simple proportional class-balancing technique, to enable better anchor matching of the OD models. A comparison was carried out on the performances of the SOTA OD models: YOLOv5, YOLOv7 and SSD, for combinations of real and synthetic datasets within an industrial use case.

Enhancing Object Detection Performance for Small Objects through Synthetic Data Generation and Proportional Class-Balancing Technique: A Comparative Study in Industrial Scenarios

TL;DR

The paper addresses the poor performance of object detectors on small industrial objects due to data scarcity and imbalance. It introduces synthetic data generation using CAD models and a proportional class-balancing scheme to balance small-object samples and improve anchor matching, evaluated on YOLOv5, YOLOv7, and SSD in an industrial assembly scenario. Across datasets, a balanced mix where synthetic data equals about half of real data (DS-3) yields the best small-object AP gains, with YOLOv5 delivering the strongest overall results; excessive synthetic data can hurt performance due to non-photo-realistic rendering. The work highlights practical trade-offs between realism and data diversity, suggesting that improved synthetic realism could further boost small-object detection in industry.

Abstract

Object Detection (OD) has proven to be a significant computer vision method in extracting localized class information and has multiple applications in the industry. Although many of the state-of-the-art (SOTA) OD models perform well on medium and large sized objects, they seem to under perform on small objects. In most of the industrial use cases, it is difficult to collect and annotate data for small objects, as it is time-consuming and prone to human errors. Additionally, those datasets are likely to be unbalanced and often result in an inefficient model convergence. To tackle this challenge, this study presents a novel approach that injects additional data points to improve the performance of the OD models. Using synthetic data generation, the difficulties in data collection and annotations for small object data points can be minimized and to create a dataset with balanced distribution. This paper discusses the effects of a simple proportional class-balancing technique, to enable better anchor matching of the OD models. A comparison was carried out on the performances of the SOTA OD models: YOLOv5, YOLOv7 and SSD, for combinations of real and synthetic datasets within an industrial use case.
Paper Structure (12 sections, 4 figures, 4 tables)

This paper contains 12 sections, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Background & Related Works
  3. Methodology
  4. Data Distribution
  5. Dataset Generation
  6. Objected Detection Models
  7. YOLOv5
  8. YOLOv7
  9. SSD
  10. Hyper-Parameters
  11. Results
  12. Conclusion

Figures (4)

  • Figure 1: Detection Leader board of COCO Dataset Object Detection Challenge from 2020. The Average Precision of Small ($AP^S$) and Large ($AP^L$) objects in the dataset for the top performing models are highlighted in the figure.
  • Figure 2: Data Distribution of Initial real dataset (left) to the combined dataset DS-3 (right). The target classes have been proportionally balanced, giving more significance for the less occurring classes.
  • Figure 3: A sample synthetic image generated using the 3D rendering Game Engine, used in the dataset.
  • Figure 4: A sample of real image data from the Assembly Scenario, taken from the live-stream video of AR device.