VISTA-3D : Training-free unfolding for vision-based 3D object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Neuromorphic Engineering

Background:

• 3D object detection is crucial for autonomous driving and AI but is computationally expensive.
• Existing neuromorphic conversion methods are not suitable for modern 3D detectors due to normalization issues.
• Neuromorphic computation offers potential for event-driven sparsity and efficiency.

Purpose of the Study:

• To develop a training-free method for adapting vision-based 3D object detectors for neuromorphic hardware.
• To address normalization-dependent inconsistencies in prior conversion techniques.
• To enable efficient and accurate 3D perception in resource-constrained AI systems.

Main Methods:

• Introduced VISTA-3D, a training-free unfolding method for 3D object detection.
• Replaced Layer Normalization (LayerNorm) with calibrated Exponential Normalization (ExpNorm).
• Created temporally unfolded representations for spike-style computation, matching one-shot ANN outputs.

Main Results:

• VISTA-3D maintained the original detector's accuracy on the KITTI benchmark (3D Average Precision, R40 protocol).
• The method generalized to larger transformer-based detectors on nuScenes, preserving competitive accuracy.
• Achieved a normalized energy proxy of ~0.17 with reduced analytic latency, without adding parameters.

Conclusions:

• VISTA-3D offers a principled, plug-and-play solution for neuromorphic-ready 3D perception.
• The calibrated unfolding approach robustly preserves accuracy, unlike direct test-time spiking.
• Provides a stable temporal representation for future fully spiking 3D detectors.