Revisiting Open World Object Detection (2201.00471v2)

Abstract: Open World Object Detection (OWOD), simulating the real dynamic world where knowledge grows continuously, attempts to detect both known and unknown classes and incrementally learn the identified unknown ones. We find that although the only previous OWOD work constructively puts forward to the OWOD definition, the experimental settings are unreasonable with the illogical benchmark, confusing metric calculation, and inappropriate method. In this paper, we rethink the OWOD experimental setting and propose five fundamental benchmark principles to guide the OWOD benchmark construction. Moreover, we design two fair evaluation protocols specific to the OWOD problem, filling the void of evaluating from the perspective of unknown classes. Furthermore, we introduce a novel and effective OWOD framework containing an auxiliary Proposal ADvisor (PAD) and a Class-specific Expelling Classifier (CEC). The non-parametric PAD could assist the RPN in identifying accurate unknown proposals without supervision, while CEC calibrates the over-confident activation boundary and filters out confusing predictions through a class-specific expelling function. Comprehensive experiments conducted on our fair benchmark demonstrate that our method outperforms other state-of-the-art object detection approaches in terms of both existing and our new metrics. Our benchmark and code are available at https://github.com/RE-OWOD/RE-OWOD.

• The paper introduces a novel framework incorporating a Proposal ADvisor and a Class-specific Expelling Classifier to improve unknown object detection.
• It establishes five benchmark principles and proposes new evaluation metrics (UDR and UDP) to realistically assess open world detection.
• Extensive experiments validate that the framework significantly boosts detection accuracy and supports incremental learning of unknown classes.

Revisiting Open World Object Detection

The paper, "Revisiting Open World Object Detection," presents an insightful examination of the Open World Object Detection (OWOD) task, a nuanced problem that attempts to bridge the gap between human intelligence and machine learning in a dynamically evolving world. The problem is characterized by a system capable of simultaneously detecting known classes while identifying and incrementally learning from unknown classes. The work critiques previous methodologies, introduces a revised experimental framework, and proposes innovative solutions to address longstanding challenges within OWOD.

Methodological Reassessment and Framework Proposal

The authors identify several weaknesses in prior OWOD research, including unsuitable benchmark configurations, unsound metric calculations, and flawed methods that compromise detection performance. To rectify these issues, they propose five fundamental benchmark principles ensuring realistic simulation, including Class Openness, Task Increment, Annotation Specificity, Label Integrity, and Data Specificity. This careful construction of benchmarks underlines the potential of OWOD to replicate real-world conditions, thus providing a credible foundation for evaluation.

The paper introduces two novel evaluation protocols, Unknown Detection Recall (UDR) and Unknown Detection Precision (UDP), aimed at quantifying detection performance from the perspective of unknown classes. These metrics address the unique challenges posed by OWOD – distinguishing unknown instances from the background and known classes – bridging a critical gap left by more conventional metrics like mean Average Precision (mAP), Wilderness Impact (WI), and Absolute Open-Set Error (A-OSE).

Enhanced OWOD Framework Components

The authors propose a novel OWOD framework incorporating an auxiliary Proposal ADvisor (PAD) and a Class-specific Expelling Classifier (CEC). The PAD module assists Region Proposal Networks (RPNs) in identifying potential unknown object proposals, enhancing detection with the integration of unsupervised object detection techniques. This module promises significant improvements through two mechanisms: confirming unknown proposals for reliable classification and guiding RPN learning to distinguish unknown proposals from the background.

Complementing the PAD, the CEC improves classification accuracy by mitigating the known problem of neural network overconfidence. The CEC expels confusing unknown predictions from known class boundaries through a class-specific refining process. This adjustment fundamentally recalibrates classification boundaries, ensuring more accurate prediction of unknown instances.

Experimental Validation

Extensive experiments demonstrate the proposed model outperforms existing state-of-the-art methods in maintaining accuracy across known classes while significantly enhancing detection capabilities for unknown classes. Notably, it shows competence in handling the incremental learning of new classes without impacting previously trained class accuracy. The ablation studies confirm the individual and combined efficacy of the PAD and CEC modules, emphasizing their roles in enriching detection precision and recall.

Practical Implications and Future Directions

The research potentially transforms OWOD from theoretical frameworks to applicable methodologies in real-world scenarios. As the framework effectively distinguishes unknown objects and incrementally incorporates them as known classes, it could redefine applications within autonomous systems, surveillance, and robotics. The clear roadmap laid out by the authors for improving OWOD benchmarks and evaluation protocols enhances its prospects of practical deployment.

Given the framework's foundation and positive results, future research might focus on refining auxiliary proposal methods and adaptive classifiers while exploring alternative network architectures to potentially heighten the OWOD performance further. The ongoing development towards more accurate models predicates substantial contributions to the field of Artificial Intelligence, aiding systems in mimicking the human ability to understand and adapt to an ever-evolving environment.