LIX: Implicitly Infusing Spatial Geometric Prior Knowledge Into Visual Semantic Segmentation for Autonomous Driving.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Data-fusion networks excel in visual semantic segmentation but require spatial geometric data.
• Transferring spatial geometric knowledge to single-modal networks is challenging but offers practical benefits.

Purpose of the Study:

• To develop a knowledge distillation framework (LIX) for infusing spatial geometric priors into single-modal networks.
• To address limitations in existing decoupled knowledge distillation methods.

Main Methods:

• Introduced the Learning to Infuse "X" (LIX) framework.
• Developed novel logit distillation with a dynamic weight controller.
• Implemented adaptively-recalibrated feature distillation using kernel regression and centered kernel alignment.

Main Results:

• The LIX framework significantly improved visual semantic segmentation performance.
• Demonstrated superior quantitative and qualitative results compared to state-of-the-art methods.
• Validated effectiveness across intermediate-fusion and late-fusion networks on public datasets.

Conclusions:

• The LIX framework effectively transfers spatial geometric knowledge, enhancing single-modal network performance.
• Novel logit and feature distillation techniques overcome existing limitations.
• The proposed methods offer a practical solution for improving visual semantic segmentation without requiring explicit spatial geometric data.