# PROB: Probabilistic Objectness for Open World Object Detection

**Authors:** Orr Zohar, Kuan-Chieh Wang, Serena Yeung  
**arXiv:** [2212.01424](https://arxiv.org/abs/2212.01424)  
**Date:** 2022-12-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.01424v1/

## Key points

- Open World Object Detection (OWOD) aims to bridge the gap between traditional object detection benchmarks and real-world scenarios
- OWOD algorithms need to detect and learn about novel or unknown objects
- Standard object detection methods fail in OWOD due to the lack of supervision in distinguishing unknown objects from background proposals
- Previous approaches using pseudo-labeling techniques have low performance in unknown object detection
- The authors propose a novel probabilistic framework called Probabilistic Objectness (PROB) for estimating the likelihood of objectness
- PROB alternates between probability distribution estimation and maximizing the objectness likelihood of known objects in an embedded feature space
- PROB integrates this probabilistic framework into traditional object detection models, adapting them for open-world settings
- Comprehensive experiments show that PROB outperforms existing methods in both unknown object detection and known object detection
- PROB's incremental learning performance is evaluated on PASCAL VOC 2007 benchmark, demonstrating its effectiveness in adapting to new objects over time
- This work presents a promising approach for open-world object detection by leveraging a probabilistic framework that improves both known and unknown object detection performance

## Summary

Open World Object Detection (OWOD) is a challenging computer vision task that aims to bridge the gap between traditional object detection benchmarks and real-world object detection scenarios. In addition to detecting and classifying known objects, OWOD algorithms are expected to detect and learn about novel or unknown objects. However, applying standard object detection methods to OWOD fails because unknown objects are often predicted as background due to the lack of supervision in distinguishing them from background object proposals. Previous approaches have attempted to address this issue by using pseudo-labeling techniques, but the performance of unknown object detection remains low. To overcome this challenge, the authors propose a novel probabilistic framework called Probabilistic Objectness (PROB) for estimating the likelihood of objectness. The framework alternates between probability distribution estimation and maximizing the objectness likelihood of known objects in an embedded feature space. This allows for estimating the objectness probability of different proposals. The resulting PROB model integrates this probabilistic framework into traditional object detection models, adapting them for open-world settings. The authors conduct comprehensive experiments on OWOD benchmarks and show that PROB outperforms existing methods in both unknown object detection (approximately 2 times higher recall) and known object detection (approximately 10% higher mean average precision). In addition, the authors evaluate PROB's incremental learning performance on the PASCAL VOC 2007 benchmark compared to other OWOD and incremental learning methods. The results demonstrate the effectiveness of PROB in adapting to new objects over time. Overall, this work presents a promising approach for addressing the challenge of open-world object detection by leveraging a probabilistic framework that improves both known and unknown object detection performance. The code for PROB will be made available upon publication at https://github.com/orrzohar/PROB.