# Revisiting Open World Object Detection

**Authors:** Xiaowei Zhao, Xianglong Liu, Yifan Shen, Yixuan Qiao, Yuqing Ma, Duorui Wang  
**arXiv:** [2201.00471](https://arxiv.org/abs/2201.00471)  
**Date:** 2022-01-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2201.00471v2/

## Key points

- Authors address challenges of Open World Object Detection (OWOD)
- Identify limitations in previous OWOD work
- Propose rethinking of experimental setting
- Introduce five fundamental benchmark principles for guiding benchmark construction
- Design two fair evaluation protocols focusing on unknown classes
- Present novel OWOD framework with Proposal Advisor (PAD) and Class-specific Expelling Classifier (CEC)
- Comprehensive experiments show proposed method outperforms other approaches in existing and new metrics
- Insights provided on improving OWOD by addressing issues with experimental settings, benchmark design, metric calculation, and methodology

## Summary

In the paper titled "Revisiting Open World Object Detection," authors Xiaowei Zhao, Xianglong Liu, Yifan Shen, Yixuan Qiao, Yuqing Ma, and Duorui Wang address the challenges of Open World Object Detection (OWOD). <br><br>

 aims to detect both known and unknown classes in a dynamic world where knowledge continuously evolves. The authors identify limitations in previous OWOD work and propose a rethinking of the experimental setting. They also introduce five fundamental benchmark principles for guiding benchmark construction. <br><br>

To address these limitations, the authors design two fair evaluation protocols that focus on evaluating from the perspective of unknown classes. These protocols aim to fill the void in existing evaluation methods. Additionally, they present a novel OWOD framework that includes an auxiliary Proposal Advisor (PAD) and a Class-specific Expelling Classifier (CEC). <br><br>

The PAD assists the Region Proposal Network (RPN) in accurately identifying unknown proposals without supervision. On the other hand, CEC calibrates over-confident activation boundaries and filters out confusing predictions using a class-specific expelling function. Comprehensive experiments conducted on their fair benchmark demonstrate that their proposed method outperforms other state-of-the-art object detection approaches in terms of both existing metrics and new metrics introduced by the authors. <br><br>

Overall, this paper provides valuable insights into improving OWOD by addressing issues with experimental settings, benchmark design, metric calculation, and methodology. The proposed benchmark principles and evaluation protocols contribute to evaluating OWOD from the perspective of unknown classes. The introduced OWOD framework with PAD and CEC demonstrates superior performance compared to existing approaches.