OW-Adapter: Human-Assisted Open-World Object Detection with a Few Examples.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Open-world object detection (OWOD) aims to identify both known and novel object classes.
• Existing deep learning models for OWOD require architectural changes, training from scratch, and extensive annotations.

Purpose of the Study:

• To develop a framework enabling pre-trained general object detectors to perform open-world object detection.
• To address challenges of existing OWOD methods, including architectural modifications, retraining, and annotation costs.

Main Methods:

• Introduced OW-Adapter, a visual analytic framework acting as an adaptor for pre-trained detectors.
• Developed a method to identify, summarize, and annotate unknown objects with minimal human effort.
• Integrated a lightweight classifier for newly annotated unknown classes into pre-trained detectors.

Main Results:

• Demonstrated the framework's effectiveness in common object recognition and autonomous driving domains.
• Showcased that OW-Adapter can extend pre-trained detectors to detect unknown objects.
• Achieved simultaneous improvement in detecting both known and unknown object classes.

Conclusions:

• OW-Adapter provides a simple and effective solution for extending general object detectors to the open-world setting.
• The framework reduces the need for extensive retraining and costly annotations for unknown classes.
• This approach enhances the versatility of existing computer vision models for diverse applications.