Presentation
Operator-Level Acceleration for Sparse 3D Object Detection
Description3D object detection is critical for robotics and autonomous driving; yet, the inference latency remains a bottleneck for efficient deployment on resource-constrained platforms. Existing acceleration strategies for sparse 3D detectors fail to fully exploit the intrinsic sparsity of voxels. To address this challenge, we propose a systematic operator-level acceleration pattern that exploits structural properties of sparse voxels through three synergistic techniques: (1) cache-aware attention fusion that redesigns slot-based aggregation via tiled tensor decomposition, achieving 10.0–33.5× speedup on GTX 4090 GPU and 12.5–50.8× speedup on A100 GPU while enabling previously infeasible high-dimensional attention heads (𝑑=512−1024); (2) sparsity-driven convolution approximation that leverages attention-induced feature redundancy to replace 3D axial convolutions with lightweight 1D channel operations, delivering 1.2–2× acceleration with negligible accuracy loss (<4% mAP); and (3) structure-aware precision reduction employing profiling-guided mixed-precision assignment that selectively applies FP16 to high-cost operators with high tolerance to precision loss, while preserving FP32 for geometric encoders, achieving 1.93–2.19× speedup. All optimizations preserve network architecture and require no retraining. Experimental results show that our method achieves a 2.1x-2.5x end-to-end speedup over state-of-the-art baselines while maintaining accuracy within 4% mAP and 2.5% NDS, establishing a novel pattern for efficient sparse 3D perception.
Event Type
Research Manuscript
TimeWednesday, July 294:54pm - 5:06pm PDT
LocationMtg Room 101A
Similar Presentations