# UC-OWOD: Unknown-Classified Open World Object Detection

**Authors:** Zhiheng Wu, Yue Lu, Xingyu Chen, Zhengxing Wu, Liwen Kang, Junzhi Yu  
**arXiv:** [2207.11455](https://arxiv.org/abs/2207.11455)  
**Date:** 2022-07-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.11455v1/

## Key points

- The paper addresses the challenging problem of Open World Object Detection (OWOD) in computer vision
- Existing methods struggle to distinguish between multiple instances of unknown classes, making accurate classification difficult
- The authors propose a novel OWOD problem called Unknown-Classified Open World Object Detection (UC-OWOD)
- UC-OWOD aims to detect and classify unknown instances into different unknown classes
- A two-stage object detector is developed, with an unknown label-aware proposal and an unknown-discriminative classification head in the first stage
- Similarity-based unknown classification and unknown clustering refinement modules are introduced in the second stage to distinguish between multiple instances of unknown classes
- Two novel evaluation protocols are introduced for assessing the detection of unknown-class objects
- Extensive experiments and visualizations demonstrate the effectiveness of the proposed method
- The availability of code on GitHub allows for replication and further research.

## Summary

, , , , 

The paper titled "UC-OWOD: Unknown-Classified Open World Object Detection" addresses the challenging problem of Open World Object Detection (OWOD) in computer vision. OWOD involves detecting unknown objects and gradually learning their identified unknown classes. Existing methods struggle to distinguish between multiple instances of unknown classes, making it difficult to accurately classify them. To overcome this limitation, the authors propose a novel OWOD problem called Unknown-Classified Open World Object Detection (UC-OWOD). The main objective of UC-OWOD is to not only detect unknown instances but also classify them into different unknown classes. To solve UC-OWOD effectively, the authors develop a two-stage object detector. In the first stage, they use an unknown label-aware proposal and an unknown-discriminative classification head to detect both known and unknown objects. This stage helps in identifying instances that belong to known classes as well as those that are part of the unknown classes. In the second stage, the authors introduce similarity-based unknown classification and unknown clustering refinement modules. These modules are designed to distinguish between multiple instances of unknown classes by leveraging similarities among them. By incorporating these modules into their framework, they improve the accuracy of classifying objects into distinct categories within the set of unknown classes. To evaluate the performance of UC-OWOD, two novel evaluation protocols are introduced specifically for assessing the detection of unknown-class objects. The authors conduct extensive experiments and provide visualizations to demonstrate the effectiveness of their proposed method. In conclusion, this paper presents a comprehensive approach for addressing OWOD through UC-OWOD. By combining a two-stage object detector with similarity-based classification and clustering refinement modules, the proposed method achieves improved detection and classification accuracy for both known and previously unidentified objects. The availability of code on GitHub allows researchers to replicate and build upon this work further.