# Large-Scale Object Detection in the Wild from Imbalanced Multi-Labels

**Authors:** Junran Peng, Xingyuan Bu, Ming Sun, Zhaoxiang Zhang, Tieniu Tan, Junjie Yan  
**arXiv:** [2005.08455](https://arxiv.org/abs/2005.08455)  
**Date:** 2020-05-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2005.08455v1/

## Key points

- The paper addresses the challenge of improving performance in deep learning through training with more data
- The authors focus on Open Images, the largest object detection dataset to date
- Open Images suffers from label-related problems such as multiple labels and imbalanced distribution
- The authors propose a concurrent softmax method for handling multi-label problems and a soft-sampling method with a hybrid training scheduler to address label imbalance
- Their approach achieves a significant improvement of 3.34 points in mean Average Precision (mAP)
- Their method achieves the best single model performance with 60.90 mAP on the public object detection test set of Open Images
- Their ensembling result achieves an even higher mAP score of 67.17, surpassing the previous best result by 4.29 points
- The paper provides valuable insights into addressing label-related challenges in large-scale object detection datasets like Open Images and demonstrates its effectiveness in improving performance and achieving state-of-the-art results

## Summary

The paper titled "Large-Scale Object Detection in the Wild from Imbalanced Multi-Labels" addresses the challenge of improving performance in deep learning through training with more data. The authors focus on Open Images, the largest object detection dataset to date, which presents both opportunities and challenges for object detection in various scenarios. Due to its semi-automatic collecting and labeling pipeline, the Open Images dataset suffers from label-related problems such as objects having multiple labels and an extremely imbalanced label distribution. To address these issues, the authors conduct a quantitative analysis and propose a simple yet effective solution. They introduce a concurrent softmax method to handle multi-label problems in object detection and a soft-sampling method with a hybrid training scheduler to address label imbalance. The results of their approach are impressive; their method achieves a significant improvement of 3.34 points leading to the best single model performance with 60.90 mAP (mean Average Precision) on the public object detection test set of Open Images. Furthermore, their ensembling result achieves an even higher mAP score of 67.17 surpassing the previous best result by 4.29 points. This paper provides valuable insights into addressing label-related challenges in large-scale object detection datasets like Open Images and demonstrates its effectiveness in improving performance and achieving state-of-the-art results in object detection tasks.