Evolving Into a Transformer: From a Training-Free Retrieval-Based Method for Anomaly Obstacle Segmentation.

Area of Science:

• Computer Vision
• Autonomous Systems
• Machine Learning

Background:

• Anomaly Obstacle Segmentation (AOS) is critical for autonomous vehicle safety, aiming to detect unforeseen obstacles.
• Existing AOS methods often require extensive retraining or image regeneration, leading to high computational costs and potential performance degradation.
• Current approaches inadequately leverage the inherent characteristics of driving scenarios.

Purpose of the Study:

• To develop a more efficient and effective Anomaly Obstacle Segmentation (AOS) method for autonomous driving.
• To reduce the computational burden and preserve the performance of semantic perceptual models.
• To improve the tolerance of perception systems to unknown objects in real-world driving conditions.

Main Methods:

• A training-free retrieval-based method is proposed, utilizing cosine similarity of appearance features to differentiate obstacles from road textures.
• The method focuses on leveraging the priors of driving scenarios to simplify the AOS task.
• A novel Transformer architecture is developed, inspired by the self-attention mechanism and the retrieval-based approach.

Main Results:

• The proposed retrieval-based method significantly outperforms existing methods in the same category by approximately 20 percentage points.
• The developed Transformer model, derived from the retrieval method, further enhances performance.
• The approach demonstrates effectiveness without requiring image re-generation or perceptual model retraining.

Conclusions:

• Paying attention to driving scenario characteristics simplifies Anomaly Obstacle Segmentation.
• The training-free retrieval method offers a computationally efficient and high-performing solution for AOS.
• The novel Transformer architecture shows promise for advancing autonomous vehicle perception capabilities.