Presentation
Sharedkd: Gradient-Guided Dynamic Student Discovery Inside a Single 3D Object Detector
DescriptionWhile state-of-the-art 3D object detectors achieve high accuracy, their computational cost hinders deployment. Knowledge Distillation (KD) offers compression, but existing methods use separate teacher-student networks, incurring memory overhead with static teachers that cannot adapt to student progress. We introduce SharedKD, unifying pruning and distillation within a single network. The full network acts as a dynamic teacher while a pruned sub-network serves as student, eliminating separate teachers, reducing memory, and enabling co-evolution for effective knowledge transfer. On nuScenes, SharedKD achieves 2.55% higher NDS than prior state-of-the-art at 75% pruning ratio, demonstrating exceptional accuracy-efficiency balance.
Event Type
Research Manuscript
TimeTuesday, July 283:54pm - 4:06pm PDT
LocationMtg Room 101A