A novel few-shot object detection framework for multi-scene driving based on contrastive proposal encoding.

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving Systems

Background:

• Few-shot object detection is crucial for autonomous systems but struggles with cross-scenario data variations and limited training examples.
• Existing methods often fail to generalize across different driving conditions (e.g., nighttime vs. daytime).

Purpose of the Study:

• To develop a robust few-shot object detection algorithm for multi-scenario driving environments.
• To address challenges of cross-scenario heterogeneity and overfitting in low-data regimes.
• To improve generalization and accuracy in diverse driving conditions.

Main Methods:

• Proposed a few-shot object detection algorithm based on FSCE, tailored for multi-scenario driving.
• Integrated a multi-scale feature module for enhanced feature representation, combining local and contextual information.
• Replaced the traditional Softmax classifier with a cosine Softmax classifier, employing L2 normalization and angular margin constraints to reduce intra-class variance.

Main Results:

• Achieved superior generalization and accuracy compared to existing methods on FLIR and BDD100K datasets.
• Demonstrated effectiveness in both nighttime infrared and daytime visible-light driving scenarios, a novel application for few-shot detection.
• Successfully addressed cross-scenario heterogeneity and overfitting in low-data regimes.

Conclusions:

• The proposed few-shot object detection algorithm offers a significant advancement for autonomous driving systems operating in varied conditions.
• The method provides a robust solution for handling data scarcity and domain shifts in real-world driving scenarios.
• Future research will focus on optimizing model complexity without compromising performance.