Reproduce “LIX: Implicitly Infusing Spatial Geometric Prior Knowledge into Visual Semantic Segmentation for Autonomous Driving” on OpenResearch

About this paper

Sicen Guo, Ziwei Long, Zhiyuan Wu, Qijun Chen, Ioannis Pitas, Rui Fan

The LIX framework introduces a knowledge distillation approach to implicitly transfer spatial geometric prior knowledge from a multi-modal teacher network (RGB-X) to a single-modal student network (RGB) for autonomous driving semantic segmentation. This method enables RGB-only networks to achieve performance comparable to their multi-modal counterparts, eliminating the need for geometric sensor data during inference.

Key insights

• The Non-Target Class Logit Distillation (NCLD) term in previous decoupled KD formulations provides zero gradients to student logit optimization, limiting its effectiveness for less confident predictions.
• Dynamically adjusting logit-wise weights based on student confidence significantly improves logit distillation compared to fixed weights.
• Kernel regression effectively aligns feature scales and distributions, and CKA provides a robust, holistic measure for feature consistency between architecturally diverse teacher and student networks.

Summary by alphaXiv.