UniDetector: Towards Universal Object Detection With Heterogeneous Supervision.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Universal object detection faces limitations due to reliance on human annotations, restricted visual data, and novel open-world categories.
• Existing detectors struggle with universality, hindering their ability to recognize all categories across diverse scenes.

Purpose of the Study:

• To propose UniDetector, a novel universal object detector capable of recognizing a vast number of categories in open-world scenarios.
• To overcome the constraints of limited data and annotations in achieving broad object detection capabilities.

Main Methods:

• Leveraging multi-source images and heterogeneous label spaces via image-text alignment for comprehensive training data.
• Employing heterogeneous supervision training to reduce dependency on fully labeled datasets.
• Utilizing a decoupling training strategy and probability calibration to enhance generalization to novel categories.

Main Results:

• UniDetector successfully detects over 7,000 categories, the largest scale reported to date, using only ~500 training classes.
• Demonstrates strong zero-shot capabilities on large-vocabulary datasets, outperforming supervised baselines by over 5% without prior image exposure.
• Achieves state-of-the-art performance on 13 diverse detection datasets using only 3% of the training data.

Conclusions:

• UniDetector effectively addresses the challenge of universal object detection in open-world settings.
• The proposed methods enable robust generalization to unseen and novel categories with remarkable efficiency.
• UniDetector represents a significant advancement in scalable and versatile object recognition systems.