Instance-Aware Knowledge Distillation for Semi-Supervised Learning of an On-Board Multi-Task Dense Prediction Model for Collision Avoidance System
Researchers have developed an instance-aware knowledge distillation framework to improve semi-supervised learning for collision avoidance systems. This method generates pseudo-labels by combining domain priors from a teacher model with instance-centric knowledge from foundation models, aiming to reduce annotation costs and computational requirements for edge deployments. The resulting lightweight student model can perform multiple dense prediction tasks in real-time, such as instance segmentation and monocular depth estimation, outperforming the larger teacher model in segmentation while maintaining performance on depth estimation. The system has been validated in a country club environment using a custom dataset and a low-cost edge device. AI
IMPACT This research could enable more efficient and capable AI-powered collision avoidance systems on edge devices, reducing development costs and improving real-time performance.