[{"data":1,"prerenderedAt":372},["ShallowReactive",2],{"method-pointnetvlad2018":3},{"method":4,"reference":56,"equipment":75,"figures":112,"results":113},{"id":5,"label":6,"shortName":7,"title":8,"year":9,"era":10,"cluster":11,"scope":12,"keyIdeaZh":13,"keyIdeaEn":14,"fulltextStatus":15,"publicationStatus":16,"recommendation":17,"constructionRelevance":18,"validationEnvironment":19,"strengths":22,"limitations":27,"sensors":35,"platform":40,"estimator":42,"association":43,"timeModel":42,"deskew":44,"loopClosure":45,"globalOptimization":46,"mapRepresentation":47,"prior":48,"outputGeometry":42,"compute":49,"codeUrl":50,"codeLicense":51,"relatedVersions":52},"pointnetvlad2018","Uy & Lee, 2018","PointNetVLAD","PointNetVLAD: Deep Point Cloud Based Retrieval for Large-Scale Place Recognition",2018,"recent","C06","place_recognition_component","PointNetVLAD 結合 PointNet 的逐點特徵與 NetVLAD 聚合層，將去除地面並下採樣為 4096 點的子地圖映射為固定長度全域描述子，以最近鄰檢索完成地點辨識；並提出 lazy triplet 與 quadruplet 損失做度量學習。作者同時以 Oxford RobotCar 與三個自建區域建立點雲檢索基準。","PointNetVLAD learns a permutation-invariant global descriptor for point-cloud submaps (PointNet + NetVLAD) trained with lazy triplet\u002Fquadruplet losses and introduces retrieval benchmarks.","full_text_reviewed","peer_reviewed_published","background","not_reported（車載都市與校園資料）",[20,21],"public_benchmark","cross_site",[23,24,25,26],"End-to-end trainable, permutation-invariant global descriptor with a proof that NetVLAD is symmetric (Sec. 4.3)","Released benchmark datasets and code (abstract)","Outperformed PointNet max-pool and ModelNet-trained PointNet baselines on all four areas, e.g. 80.31% vs 73.44% and 46.52% top-1% recall on Oxford (Table 2)","Authors report point-cloud retrieval more robust than image NetVLAD for day-to-night queries (Sec. 5.2, Fig. 6c)",[28,29,30,31,32,33,34],"Requires supervised training with GPS\u002FINS-referenced maps (Sec. 5.1)","(inference) Fixed 4096-point, ground-removed, normalized submaps discard fine geometry and absolute scale cues","Authors show failures on continuous roads with very similar features and in heavily occluded areas (Sec. 5.2, Fig. 7)","(inference) Coarse success criterion: retrieval counted correct within 25 m (Sec. 5.1)","Generalization from Oxford-only training was limited (60.27% top-1% recall on R.A.) and improved after refinement on U.S. and R.A. (Tables 2 and 5)","Found in follow-up work: with released weights PointNetVLAD failed under larger rotational and lateral variation on MulRan Riverside 02 and KITTI 08 (scancontextpp2022 Sec. VIII-E, Fig. 21)","Found in follow-up survey: learning-based global localization faces generalization problems across LiDAR types, mounting and unseen environments (yin2024survey Sec. 6.6)",[36,37,38,39],"2D LiDAR SICK LMS-151 scans accumulated into 3D reference maps and submaps using GPS\u002FINS (Oxford RobotCar benchmark)","Velodyne-64 LiDAR, as written (in-house U.S., R.A., B.D. sets)","GPS\u002FINS (reference maps in UTM frame and training labels)","stereo camera centre images used only for the NetVLAD image baseline",[41],"vehicle","not_applicable","learned global descriptor (PointNet features + NetVLAD aggregation + fully connected layer), nearest-neighbour retrieval; lazy triplet\u002Fquadruplet metric-learning loss","not_reported","retrieval of structurally similar submaps; no metric pose output","none (component)","database of fixed-size (4096-point) ground-removed, normalized submaps","supervised training with geo-referenced submaps (positives within 10 m, negatives beyond 50 m)","TensorFlow implementation; inference about 9 ms on an NVIDIA GeForce GTX 1080Ti; retrieval through the submap database O(log n); training with batches of 3 tuples, 18 hard negatives each","https:\u002F\u002Fgithub.com\u002Fmikacuy\u002Fpointnetvlad","MIT (LICENSE file)",[53],{"relation":54,"title":55,"doi_or_url":50},"code_release","mikacuy\u002Fpointnetvlad",{"id":5,"kind":57,"shortName":7,"title":8,"authors":58,"year":9,"venue":61,"venueType":62,"publisher":63,"volumeIssuePages":64,"doi":65,"arxivId":66,"url":67,"firstPublicDate":68,"publicationStatus":16,"metadataStatus":69,"fulltextStatus":15,"era":10,"classicReason":42,"codeUrl":50,"cluster":11,"topics":70,"mdpi":71,"verification":72,"label":6,"fulltextRoute":73,"versionRead":74,"addedByCensus":71},"component",[59,60],"Mikaela Angelina Uy","Gim Hee Lee","2018 IEEE\u002FCVF Conference on Computer Vision and Pattern Recognition (CVPR)","conference","IEEE","pp. 4470-4479","10.1109\u002Fcvpr.2018.00470","1804.03492","https:\u002F\u002Farxiv.org\u002Fabs\u002F1804.03492","2018-04-10","metadata_verified",[11],false,"confirmed","arXiv","arXiv 1804.03492v3 (16 May 2018), CVPR 2018 paper with supplementary material (11 pages); CVF or IEEE version not compared",[76,83,90,95,99,104],{"category":77,"model":78,"canonical":78,"role":79,"dataset":80,"specs":81,"locator":82},"lidar","SICK LMS-151","dataset sensor","Oxford RobotCar","2D LiDAR scanner; 44 full and partial runs","Sec. 5.1 Oxford Dataset",{"category":84,"model":85,"canonical":85,"role":86,"dataset":87,"specs":88,"locator":89},"gnss","GPS\u002FINS (model not stated)","reference or ground truth","Oxford RobotCar; in-house U.S., R.A., B.D.","reference maps built in UTM coordinates","Sec. 5.1",{"category":77,"model":91,"canonical":91,"role":79,"dataset":92,"specs":93,"locator":94},"Velodyne-64 LiDAR","in-house U.S., R.A., B.D.","mounted on a car; five runs per region","Sec. 5.1 In-house Datasets",{"category":96,"model":97,"canonical":97,"role":79,"dataset":80,"specs":44,"locator":98},"stereo_camera","centre stereo camera of the Oxford RobotCar platform (model not stated)","Sec. 5.2 Image based comparisons",{"category":100,"model":101,"canonical":101,"role":79,"dataset":102,"specs":103,"locator":89},"platform","car","Oxford RobotCar; in-house sets","routes of 10, 10, 8 and 5 km per round for Oxford, U.S., R.A., B.D.",{"category":105,"model":106,"canonical":107,"role":108,"dataset":109,"specs":110,"locator":111},"compute","NVIDIA GeForce GTX 1080Ti","Nvidia GeForce GTX1080 Ti","compute for runtime",null,"about 9 ms inference","Sec. 5.2 Usability",[],{"totalRows":114,"groupCount":115,"groups":116,"others":360},32,6,[117,217,293,329],{"slug":118,"group":119,"sourceId":5,"sourceLabel":6,"table":120,"selfRows":121,"metrics":122,"seqs":127,"entrants":141,"cells":155,"outcomes":211,"locators":212,"hardware":213,"wordings":214,"notes":215},"pointnetvlad2018-table-3","pointnetvlad2018:Table 3","Table 3",12,[123],{"label":124,"unit":125,"statistic":126,"alignment":42},"average recall (%) at top 1%","%","mean",[128,132,135,138],{"dataset":129,"sequence":130,"environment":131},"Oxford RobotCar benchmark","test reference maps","not_reported (Oxford RobotCar routes; environment type not stated in the paper)",{"dataset":133,"sequence":130,"environment":134},"In-house U.S.","university sector",{"dataset":136,"sequence":130,"environment":137},"In-house R.A.","residential area",{"dataset":139,"sequence":130,"environment":140},"In-house B.D.","business district",[142,145,147,149,151,153],{"name":143,"methodId":5,"linkable":144,"proposed":144,"self":144},"PN_VLAD (PointNetVLAD) D-128",true,{"name":146,"methodId":5,"linkable":144,"proposed":144,"self":144},"PN_VLAD (PointNetVLAD) D-256",{"name":148,"methodId":5,"linkable":144,"proposed":144,"self":144},"PN_VLAD (PointNetVLAD) D-512",{"name":150,"methodId":109,"linkable":71,"proposed":71,"self":71},"PN_MAX (PointNet + maxpool + FC) D-128",{"name":152,"methodId":109,"linkable":71,"proposed":71,"self":71},"PN_MAX (PointNet + maxpool + FC) D-256",{"name":154,"methodId":109,"linkable":71,"proposed":71,"self":71},"PN_MAX (PointNet + maxpool + FC) D-512",[156,160,163,166,169,172,175,177,179,181,183,185,187,189,191,193,195,197,199,201,203,205,207,209],[157,157,157,158,159,157,159,159,157],0,74.6,-1,[161,157,157,162,159,157,159,159,157],1,80.31,[164,157,157,165,159,157,159,159,157],2,80.33,[167,157,157,168,159,157,159,159,157],3,71.93,[170,157,157,171,159,157,159,159,157],4,73.44,[173,157,157,174,159,157,159,159,157],5,74.79,[157,157,161,176,159,157,159,159,157],66.03,[161,157,161,178,159,157,159,159,157],72.63,[164,157,161,180,159,157,159,159,157],76.24,[167,157,161,182,159,157,159,159,157],61.15,[170,157,161,184,159,157,159,159,157],64.64,[173,157,161,186,159,157,159,159,157],65.79,[157,157,164,188,159,157,159,159,157],53.86,[161,157,164,190,159,157,159,159,157],60.27,[164,157,164,192,159,157,159,159,157],63.31,[167,157,164,194,159,157,159,159,157],49.25,[170,157,164,196,159,157,159,159,157],51.92,[173,157,164,198,159,157,159,159,157],52.32,[157,157,167,200,159,157,159,159,157],59.84,[161,157,167,202,159,157,159,159,157],65.3,[164,157,167,204,159,157,159,159,157],66.75,[167,157,167,206,159,157,159,159,157],53.25,[170,157,167,208,159,157,159,159,157],54.74,[173,157,167,210,159,157,159,159,157],56.63,[],[120],[],[],[216],"Output dimensionality D of the global descriptor; trained on Oxford; average recall at top 1%",{"slug":218,"group":219,"sourceId":5,"sourceLabel":6,"table":220,"selfRows":221,"metrics":222,"seqs":226,"entrants":231,"cells":238,"outcomes":287,"locators":288,"hardware":289,"wordings":290,"notes":291},"pointnetvlad2018-table-5","pointnetvlad2018:Table 5","Table 5",8,[223,224],{"label":124,"unit":125,"statistic":126,"alignment":42},{"label":225,"unit":125,"statistic":126,"alignment":42},"average recall (%) at top 1",[227,228,229,230],{"dataset":129,"sequence":130,"environment":131},{"dataset":133,"sequence":130,"environment":134},{"dataset":136,"sequence":130,"environment":137},{"dataset":139,"sequence":130,"environment":140},[232,234,236],{"name":233,"methodId":5,"linkable":144,"proposed":144,"self":144},"PN_VLAD (PointNetVLAD)",{"name":235,"methodId":109,"linkable":71,"proposed":71,"self":71},"PN_MAX (PointNet + maxpool + FC)",{"name":237,"methodId":109,"linkable":71,"proposed":71,"self":71},"PN_STD (PointNet trained on ModelNet)",[239,241,243,245,247,249,251,253,255,257,259,261,263,265,267,269,271,273,275,277,279,281,283,285],[157,157,157,240,159,157,159,159,157],80.09,[161,157,157,242,159,157,159,159,157],73.87,[164,157,157,244,159,157,159,159,157],46.52,[157,161,157,246,159,157,159,159,157],63.33,[161,161,157,248,159,157,159,159,157],54.16,[164,161,157,250,159,157,159,159,157],31.87,[157,157,161,252,159,157,159,159,157],90.1,[161,157,161,254,159,157,159,159,157],79.31,[164,157,161,256,159,157,159,159,157],56.95,[157,161,161,258,159,157,159,159,157],86.07,[161,161,161,260,159,157,159,159,157],62.16,[164,161,161,262,159,157,159,159,157],45.67,[157,157,164,264,159,157,159,159,157],93.07,[161,157,164,266,159,157,159,159,157],75.14,[164,157,164,268,159,157,159,159,157],59.81,[157,161,164,270,159,157,159,159,157],82.66,[161,161,164,272,159,157,159,159,157],60.21,[164,161,164,274,159,157,159,159,157],44.29,[157,157,167,276,159,157,159,159,157],86.49,[161,157,167,278,159,157,159,159,157],69.49,[164,157,167,280,159,157,159,159,157],53.02,[157,161,167,282,159,157,159,159,157],80.11,[161,161,167,284,159,157,159,159,157],58.95,[164,161,167,286,159,157,159,159,157],44.54,[],[220],[],[],[292],"Refined networks trained on Oxford, U.S. and R.A. (B.D. unseen); average recall at top 1% and at top 1",{"slug":294,"group":295,"sourceId":5,"sourceLabel":6,"table":296,"selfRows":170,"metrics":297,"seqs":299,"entrants":304,"cells":308,"outcomes":323,"locators":324,"hardware":325,"wordings":326,"notes":327},"pointnetvlad2018-table-2","pointnetvlad2018:Table 2","Table 2",[298],{"label":124,"unit":125,"statistic":126,"alignment":42},[300,301,302,303],{"dataset":129,"sequence":130,"environment":131},{"dataset":133,"sequence":130,"environment":134},{"dataset":136,"sequence":130,"environment":137},{"dataset":139,"sequence":130,"environment":140},[305,306,307],{"name":233,"methodId":5,"linkable":144,"proposed":144,"self":144},{"name":235,"methodId":109,"linkable":71,"proposed":71,"self":71},{"name":237,"methodId":109,"linkable":71,"proposed":71,"self":71},[309,310,311,312,313,314,316,317,318,320,321,322],[157,157,157,162,159,157,159,159,157],[161,157,157,171,159,157,159,159,157],[164,157,157,244,159,157,159,159,157],[157,157,161,178,159,157,159,159,157],[161,157,161,184,159,157,159,159,157],[164,157,161,315,159,157,159,159,157],61.12,[157,157,164,190,159,157,159,159,157],[161,157,164,196,159,157,159,159,157],[164,157,164,319,159,157,159,159,157],49.07,[157,157,167,202,159,157,159,159,157],[161,157,167,208,159,157,159,159,157],[164,157,167,280,159,157,159,159,157],[],[296],[],[],[328],"Baseline networks trained on Oxford only; average recall at top 1%; success if retrieved submap within 25 m",{"slug":330,"group":331,"sourceId":5,"sourceLabel":6,"table":332,"selfRows":170,"metrics":333,"seqs":335,"entrants":337,"cells":346,"outcomes":354,"locators":355,"hardware":356,"wordings":357,"notes":358},"pointnetvlad2018-table-4","pointnetvlad2018:Table 4","Table 4",[334],{"label":124,"unit":125,"statistic":126,"alignment":42},[336],{"dataset":129,"sequence":130,"environment":131},[338,340,342,344],{"name":339,"methodId":5,"linkable":144,"proposed":144,"self":144},"PN_VLAD with Triplet Loss",{"name":341,"methodId":5,"linkable":144,"proposed":144,"self":144},"PN_VLAD with Quadruplet Loss",{"name":343,"methodId":5,"linkable":144,"proposed":144,"self":144},"PN_VLAD with Lazy Triplet Loss",{"name":345,"methodId":5,"linkable":144,"proposed":144,"self":144},"PN_VLAD with Lazy Quadruplet Loss",[347,349,351,353],[157,157,157,348,159,157,159,159,157],71.2,[161,157,157,350,159,157,159,159,157],74.13,[164,157,157,352,159,157,159,159,157],78.99,[167,157,157,162,159,157,159,159,157],[],[332],[],[],[359],"PN_VLAD trained and tested on Oxford with different losses; average recall at top 1%",[361,367],{"group":362,"slug":363,"sourceLabel":364,"table":365,"selfRows":167,"datasets":366},"scancontextpp2022:Table V","scancontextpp2022-table-v","Kim et al., 2022b","Table V",[44],{"group":368,"slug":369,"sourceLabel":6,"table":370,"selfRows":161,"datasets":371},"pointnetvlad2018:Text Sec.5.2","pointnetvlad2018-text-sec-5-2","Text Sec.5.2",[44],1790510663324]