Kookmin Team Introduces SharedKD for 3D Detection
At DAC 2026 on July 28, Kookmin University researchers presented SharedKD, a model-compression method for 3D object detectors used in autonomous-driving perception. The method combines pruning and knowledge distillation in one detector, using the full network as a dynamic teacher and a pruned sub-network as the student, and the presentation reports 2.55% higher NDS than the prior state of the art on nuScenes at a 75% pruning ratio.
Kookmin University researchers introduced SharedKD, a compression approach for 3D object detection models that combines pruning and knowledge distillation within a single detector. The work, titled "SharedKD: Gradient-Guided Dynamic Student Discovery Inside a Single 3D Object Detector," appeared as a research manuscript presentation at DAC 2026 on July 28.
According to the DAC conference presentation, conventional knowledge-distillation approaches for 3D detection use separate teacher and student networks. SharedKD instead uses the full detector as a dynamic teacher while a pruned sub-network operates as the student. The presentation describes this design as eliminating the separate teacher network, reducing memory overhead, and allowing the teacher-student relationship to adapt as training progresses.
The Kookmin University AI and Software Graduate School lists the paper's authors as H. Cho, S. An, H. Park, and Jangho Kim. Kim is an assistant professor in Kookmin University's School of AI, where the Machine Learning and Pattern Recognition Lab identifies pruning, quantization, and knowledge distillation among its research areas.
Reported benchmark result
The DAC presentation reports that SharedKD achieved 2.55% higher NDS than the previous state of the art on the nuScenes benchmark at a 75% pruning ratio. NDS, or nuScenes Detection Score, combines detection quality measures for autonomous-driving perception, making it a useful aggregate metric for evaluating 3D detectors.
The reported result concerns a specific pruning setting and benchmark comparison. The available sources do not provide latency, throughput, hardware-power, or memory measurements for production deployments, so the extent of runtime benefit on an in-vehicle platform is not established by the cited materials.
Why the architecture matters
For ML practitioners, the central technical distinction is the use of an in-model teacher rather than a separately maintained full teacher network. In conventional distillation pipelines, the teacher may add training-time memory and engineering complexity, especially for high-cost 3D perception models. The conference description characterizes SharedKD's teacher as dynamic because it co-evolves with the pruned student during training.
More broadly, research on compression for autonomous-driving perception commonly has to preserve detection quality while reducing compute and memory demands. A single-network approach can be relevant where teams are evaluating pruning and distillation jointly, although deployment decisions still require measurements across target accelerators, sensor configurations, latency budgets, and safety-validation datasets.
Key Points
- 1SharedKD combines pruning and distillation inside one 3D detector, avoiding a separate teacher network during its reported training approach.
- 2The DAC presentation reports a 2.55% NDS improvement over prior work at 75% pruning on nuScenes.
- 3Single-network compression methods can reduce training-system complexity, but production relevance depends on unreported hardware latency, memory, and safety-validation results.
Scoring Rationale
SharedKD reports a notable benchmark result for efficient 3D perception, an important workload in autonomous driving and robotics. Its practitioner impact is constrained by the limited available deployment evidence, including no cited latency or hardware measurements.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
