Research map / Detection and segmentation / 3D detection and driving
FlashBEV: Fast and Memory-Efficient Exact BEV Transformation with IO-Awareness
Shunsuke Yokokawa, Hironori Kasahara
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace tensorized execution with fused FlashBEV kernel | Tensorized Sampling-VT · VT operator reference config (X=Y=200, Z=8, C=128) | forward peak memory (MB): 1971.94 -> 52.82 | Improved |
| fuse VT into one kernel | Tensorized Sampling-VT · VT operator reference config (X=Y=200, Z=8, C=128) | forward latency (ms): 9.18 -> 1.77 | Improved |
| fuse VT into one kernel | Tensorized Sampling-VT · VT operator reference config (X=Y=200, Z=8, C=128) | backward latency (ms): 17.90 -> 3.27 | Improved |
| fuse VT into one kernel | Tensorized Sampling-VT · VT operator reference config (X=Y=200, Z=8, C=128) | backward peak memory (MB): 1174.86 -> 105.63 | Improved |
| use FlashBEV as VT module end to end | SimpleBEV pipeline · E2E reference config (X=Y=200, Z=8, C=128) | E2E peak memory (MB): 2211.29 -> 876.76 | Improved |
| use FlashBEV as VT module end to end | SimpleBEV pipeline · E2E reference config (X=Y=200, Z=8, C=128) | E2E latency (ms): 67.19 -> 59.91 | Improved |
| use FlashBEV instead of tensorized VT | SimpleBEV · nuScenes BEV detection | IoU: 46.9 -> 46.8 | No clear effect |
| replace SimpleBEV height conv with height sum + Conv2d (ablation) | SimpleBEV · nuScenes BEV detection | IoU: 46.9 -> 46.9 | No clear effect |
| increase height bins Z from 8 to 32 in TensorRT (ablation) | FlashBEV plugin · TensorRT plugin | peak memory (MB): 38 -> 38 | No clear effect |
| increase height bins Z from 8 to 32 in TensorRT (ablation) | Tensorized Sampling-VT · TensorRT plugin | peak memory (MB): 1294 -> 4994 | Hurt |
| apply FlashSCA execution to BEVFormer SCA | BEVFormerSCA · BEVFormer SCA TensorRT (P=32) | peak memory (MB): 3708 -> 1168 | Improved |
| apply FlashSCA execution to BEVFormer SCA | BEVFormerSCA · BEVFormer SCA TensorRT (P=32) | latency (ms): 29.27 -> 13.90 | Improved |
Closest papers by meaning
- PointBeV: A Sparse Approach for BeV Predictions (CVPR 2024 · similarity 0.87)
- M-BEV: Masked BEV Perception for Robust Autonomous Driving (AAAI 2024 · similarity 0.85)
- BEVDiffuser: Plug-and-Play Diffusion Model for BEV Denoising with Ground-Truth Guidance (CVPR 2025 · similarity 0.85)
- FlashMask: Efficient and Rich Mask Extension of FlashAttention (ICLR 2025 · similarity 0.85)
- FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision (NeurIPS 2024 · similarity 0.85)
- DiffBEV: Conditional Diffusion Model for Bird’s Eye View Perception (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
