FlashBEV: Fast and Memory-Efficient Exact BEV Transformation with IO-Awareness

2026-07-10

AI Summary

This work addresses the scalability limitations of sampling-based BEV view transformation, which incurs prohibitive memory and latency overheads during deployment due to explicit generation of large intermediate tensors whose size scales with vertical resolution and the number of cameras. To overcome this bottleneck, the authors propose an I/O-aware operator fusion strategy that reformulates view transformation as a query-driven gathering process. By employing a gather-reduction pattern, thread-local accumulation, and on-the-fly recomputation, the method avoids materializing intermediate tensors while preserving identical output. This eliminates the memory bottleneck associated with the height dimension entirely. As a result, peak GPU memory consumption is reduced by over an order of magnitude, inference latency is significantly lowered, and memory usage becomes dependent solely on the BEV output size—enabling higher-resolution, longer-range, and finer-grained vertical discretization within a fixed hardware budget.

Abstract

Bird's-eye-view (BEV) perception is a core component of camera-based 3D understanding in autonomous driving, where view transformation (VT) maps multi-camera image features into a unified BEV representation. Sampling-based view transformation (Sampling-VT) is attractive because it supports dense and continuous BEV aggregation for high-resolution and long-range perception. Its deployment bottleneck, however, is systems-level: standard tensorized implementations of Sampling-VT -- which we refer to as Tensorized Sampling-VT -- explicitly materialize large height-dependent intermediate tensors, causing memory and latency costs that scale poorly with vertical resolution and the number of cameras. We revisit Tensorized Sampling-VT from an operator-execution perspective and show that it follows a gather-reduction pattern: each BEV query independently accumulates contributions across cameras and height bins, enabling thread-local accumulation with on-the-fly recomputation that eliminates the need to materialize height- and camera-dependent intermediates. Based on this insight, we propose FlashBEV, a fully fused and IO-aware execution strategy mathematically equivalent to Tensorized Sampling-VT (same operator output) while substantially reducing global memory traffic and kernel-launch overhead. Experiments show that FlashBEV achieves more than an order of magnitude lower peak GPU memory and significant inference-latency speedups, with memory effectively independent of the number of height bins, reducing the operator's peak memory to O(BCXY) (output only). This unlocks higher BEV range/resolution and vertical discretization within fixed deployment budgets on memory-constrained devices. Our contribution is an execution redesign -- same math, different execution -- that removes a key scalability barrier for deployment-ready Sampling-VT. Code available at https://github.com/yokosyun/FlashBEV

Problem

• Bird's-eye-view (BEV)
• View Transformation
• Memory Efficiency
• Autonomous Driving
• Sampling-based VT

Innovation

• BEV perception
• view transformation
• memory efficiency
• IO-aware execution
• tensor fusion