Track Everything: Limiting Prior Knowledge in Online Multi-Object Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Online tracking and classification of multiple objects in image sequences present significant challenges.
• Existing systems often rely on object-specific prior knowledge (hand-crafted features, user initialization), limiting generality.

Purpose of the Study:

• To develop a robust, general-purpose object recognition system for detecting and tracking diverse object types.
• To improve object recognition by integrating tracking information with a fast-learning image classifier.

Main Methods:

• Objects are tracked first without relying on object-specific prior knowledge.
• A shallow convolutional neural network (CNN) is used for fast object classification.
• Biologically inspired implementation adaptively learns object shape and motion.

Main Results:

• Object recognition performance improves when classification is combined with tracking state information.
• The proposed approach is competitive with state-of-the-art systems that use object-specific prior knowledge.
• The system demonstrates generality across multiple object types on the Neovision2 Tower benchmark.

Conclusions:

• Transferring prior knowledge from tracking to classification enables a more robust and general object recognition system.
• The method offers practical advantages due to its generality, reducing reliance on specific object features.
• The approach provides a competitive and adaptable solution for real-world video object recognition tasks.