[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85802-en":3,"doc-seo-85802-105":29,"detail-sidebar-cat-0-en-105":90},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},85802,8796095462418,"Noah","https://ap-avatar.wpscdn.com/avatar/80000253c1241d02b47?x-image-process=image/resize,m_fixed,w_180,h_180&k=1778826106357471780",6,"Technology","FlashBEV Fast and Memory-Efficient Exact BEV Transformation with IO-Awareness","Bird’s-eye-view (BEV) perception is central to camera-based 3D understanding in autonomous driving, where view transformation (VT) maps multi-camera image features into a unified BEV representation. Sampling-based VT supports dense, continuous aggregation but standard tensorized implementations materialize large height-dependent intermediates, creating memory and latency costs. By analyzing the operator as a gather–reduction pattern, FlashBEV proposes a fully fused, IO-aware execution that is mathematically equivalent to the tensorized baseline. Experiments show over-order-of-magnitude lower peak GPU memory and faster inference, enabling higher BEV range and resolution under fixed deployment budgets.","arXiv :2607 . 10071v1 [ cs .CV] 11 Jul 2026  \nFlashBEV: Fast and Memory-Efficient Exact BEV Transformation with IO-Awareness  \nShunsuke Yokokawa 1 ,2 and Hironori Kasahara 1  \n1 Waseda University, Tokyo, Japan  \n[yokosyun@fuji.waseda.jp](yokosyun@fuji.waseda.jp) , [kasahara@waseda.jp](kasahara@waseda.jp)  \n2 T2, Inc. , Tokyo, Japan  \nAbstract. 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.  \nWe 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. Unlike splatting-based VT, which requires index sorting and prevents fully thread-local reduction from voxel construction to BEV output, the gather–reduction structure of Sampling-VT enables thread-local accumulation with on-the-fly recomputation, eliminating the need to materialize height-and camera-dependent intermediates.  \nBased on this insight, we propose FlashBEV, a fully fused and IOaware execution strategy that is 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 usage 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, where tensorized execution would otherwise be infeasible. Our contribution is therefore an execution redesign—same math, different execution—that removes a keyscalability barrier for deployment-ready Sampling-VT. Code is available at [https://github.com/yokosyun/FlashBEV](https://github.com/yokosyun/FlashBEV).  \nKeywords: Bird’s-eye-view · View transformation · Memory efficiency · Autonomous driving · GPU optimization  \n2 S. Yokokawa and H. Kasahara  \nFig. 1: Conceptual comparison of Splatting-VT (left), which forward-projects (scatters) image features into BEV, and Sampling-VT (right), which backward-queries (gathers) features at each BEV location.  \n1 Introduction  \nBird’s-eye-view (BEV) perception is a fundamental representation for camerabased 3D scene understanding in autonomous driving. A central operator in BEV pipelines is view transformation (VT), which maps image-space features into a shared BEV coordinate frame. Because VT is executed on every frame, its computational efficiency directly affects end-to-end deployability.  \nExisting VT designs are broadly divided into splatting-based VT (SplattingVT) and sampling-based VT (Sampling-VT) . As shown in Fig. 1, Splatting-VT forward-projects image features into BEV and is often memory-friendly with index-based pooling [11], but it tends to produce geometry-dependent, uneven coverage. Sampling-VT instead performs backward querying from each BEV location into image space, enabling dense and continuous aggregation over the BEV grid. This density makes Sampling-VT attractive for high-resolution and long-range perception.  \nThe challenge is that Sampling-VT is expensive under standard tensorized execution. In this commonly used baseline (Tensorized Sampling-VT, i.e., tensorized","cbCaiuGcMDCO1jgk","https://ap.wps.com/l/cbCaiuGcMDCO1jgk","pdf",586422,1,17,"English","en",105,"# Introduction\n## View transformation for BEV perception\n## Splatting-based vs sampling-based VT\n## Bottleneck in tensorized Sampling-VT\n# FlashBEV execution redesign\n## Gather–reduction operator analysis\n## Fused IO-aware implementation\n## Contributions and results","[{\"question\":\"What problem does FlashBEV address in sampling-based view transformation?\",\"answer\":\"Standard tensorized Sampling-VT explicitly materializes large height- and camera-dependent intermediate tensors, causing memory and latency costs that grow poorly with vertical resolution and the number of cameras.\"},{\"question\":\"How does FlashBEV achieve exact equivalence to Tensorized Sampling-VT?\",\"answer\":\"FlashBEV redesigns the GPU execution by fusing sampling and accumulation into a thread-local gather–reduction with on-the-fly recomputation, removing intermediate tensor materialization while preserving the same operator output.\"},{\"question\":\"What performance improvements does FlashBEV report?\",\"answer\":\"Experiments show more than an order of magnitude lower peak GPU memory and significant inference-latency speedups, with peak memory effectively independent of the number of height bins.\"}]",1784206357,43,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":27},"flashbev-fast-and-memory-efficient-exact-bev-transformation-with-io-awareness","",{"@graph":35,"@context":84},[36,53,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/technology/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/flashbev-fast-and-memory-efficient-exact-bev-transformation-with-io-awareness/85802/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":61,"encodingFormat":60,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":4},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What problem does FlashBEV address in sampling-based view transformation?","Question",{"text":74,"@type":75},"Standard tensorized Sampling-VT explicitly materializes large height- and camera-dependent intermediate tensors, causing memory and latency costs that grow poorly with vertical resolution and the number of cameras.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does FlashBEV achieve exact equivalence to Tensorized Sampling-VT?",{"text":79,"@type":75},"FlashBEV redesigns the GPU execution by fusing sampling and accumulation into a thread-local gather–reduction with on-the-fly recomputation, removing intermediate tensor materialization while preserving the same operator output.",{"name":81,"@type":72,"acceptedAnswer":82},"What performance improvements does FlashBEV report?",{"text":83,"@type":75},"Experiments show more than an order of magnitude lower peak GPU memory and significant inference-latency speedups, with peak memory effectively independent of the number of height bins.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,112,117,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":110,"slug":111},50,"technology",{"id":113,"doc_module":4,"doc_module_name":45,"category_name":114,"show_sort_weight":115,"slug":116},7,"Healthcare",40,"healthcare",{"id":118,"doc_module":4,"doc_module_name":45,"category_name":119,"show_sort_weight":120,"slug":121},8,"Research & Report",30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":105,"slug":137},19,"General","general"]