[{"data":1,"prerenderedAt":403},["ShallowReactive",2],{"method-landry2019cello3d":3},{"method":4,"reference":48,"equipment":67,"figures":74,"results":75},{"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":23,"limitations":27,"sensors":32,"platform":34,"estimator":36,"association":37,"timeModel":38,"deskew":18,"loopClosure":38,"globalOptimization":39,"mapRepresentation":38,"prior":39,"outputGeometry":40,"compute":41,"codeUrl":42,"codeLicense":43,"relatedVersions":44},"landry2019cello3d","Landry et al., 2019","CELLO-3D","CELLO-3D: Estimating the Covariance of ICP in the Real World",2019,"recent","C13","registration_component","作者先檢視既有封閉形式共變異數估計在 3D 資料上的限制，再以資料驅動方式學習 ICP 配準的共變異數。訓練與評估使用超過五百萬次配準、1020 組真實點雲對，涵蓋結構化與非結構化、室內與室外環境。","Learns ICP registration covariance from large-scale real registrations after showing the limits of closed-form estimators on 3D data.","full_text_reviewed","peer_reviewed_published","supplementary","not_reported",[20,21,22],"public_benchmark","completed_building","independent_reference",[24,25,26],"Lower average KL divergence than the mean-covariance baseline on all seven test groups, with the largest gains indoors, e.g. Apartment 26.6 vs 34.1 (Table II)","Censi's closed-form estimate had average KL divergences of 2.25e6 to 2.06e8 because it was orders of magnitude too small (Table II, Sec. VI-A)","Average Mahalanobis distance of final odometry poses between 0.405 and 2.50 over 100 trajectories per dataset, which the authors judge consistent overall (Table III, Sec. VI-B)",[28,29,30,31],"Gains over the baseline are modest in self-similar environments (Wood, Gazebo), where the learned weights become nearly uniform (Sec. VI-A)","Estimates are pessimistic for Stairs and Apartment (average DM below 1.5) (Sec. VI-B)","Treating ICP results as normally distributed is error-prone in SE(3) because observed distributions are mainly multimodal, and descriptor quality is critical (Sec. VII)","Data augmentation only rotates about the z axis (2.5D), and each test group was trained on another dataset of the same environment type (Sec. V-A, VI, Table II)",[33],"3D point clouds from the 'Challenging data sets for point cloud registration algorithms' (Pomerleau et al. 2012); the sensor is not named in this paper",[35],"offline evaluation on recorded public datasets (no platform named in this paper)","CELLO-style kernel-weighted average of training covariances; an upper-triangular distance metric is learned by SGD with a determinant-plus-trace loss; descriptors per cell of a 4x4x4 grid (25 m x 25 m x 10 m) over the overlap region hold planarity, cylindricality and a 9-bin normal histogram; training covariances come from 5000 ICP samples per pair filtered with DBSCAN","point-to-plane ICP configured in the framework of Pomerleau et al. [24]: maximum-density and random subsampling filters, k-d tree 3-nearest-neighbour matching, trimmed-distance outlier filter keeping the closest 70%, at most 80 iterations","not_applicable","none","registration covariance","training covariances from about 5,100,000 registrations on 1020 pairs computed on Compute Canada clusters in about 5 CPU-years; online inference time not reported",null,"not_verified",[45],{"relation":46,"title":8,"doi_or_url":47},"preprint","https:\u002F\u002Farxiv.org\u002Fabs\u002F1810.01470",{"id":5,"kind":49,"shortName":7,"title":8,"authors":50,"year":9,"venue":54,"venueType":55,"publisher":56,"volumeIssuePages":57,"doi":58,"arxivId":59,"url":47,"firstPublicDate":60,"publicationStatus":16,"metadataStatus":61,"fulltextStatus":15,"era":10,"classicReason":38,"codeUrl":42,"cluster":11,"topics":62,"mdpi":63,"verification":64,"label":6,"fulltextRoute":65,"versionRead":66,"addedByCensus":63},"component",[51,52,53],"David Landry","Francois Pomerleau","Philippe Giguere","2019 International Conference on Robotics and Automation (ICRA)","conference","IEEE","pp. 8190-8196","10.1109\u002Ficra.2019.8793516","1810.01470","2018-10-02","metadata_verified",[11],false,"confirmed","arXiv","arXiv v1 (2018-10-02), the only arXiv version; ICRA 2019 VoR not compared",[68],{"category":69,"model":70,"canonical":70,"role":71,"dataset":42,"specs":72,"locator":73},"compute","Compute Canada computing clusters","compute for runtime","about 5 CPU-years for about 5,100,000 training registrations (offline data generation, not online runtime)","Sec. V-A",[],{"totalRows":76,"groupCount":77,"groups":78,"others":402},58,3,[79,208,322],{"slug":80,"group":81,"sourceId":5,"sourceLabel":6,"table":82,"selfRows":83,"metrics":84,"seqs":99,"entrants":120,"cells":125,"outcomes":201,"locators":202,"hardware":203,"wordings":204,"notes":205},"landry2019cello3d-table-iii","landry2019cello3d:Table III","Table III",35,[85,89,92,95,97],{"label":86,"unit":87,"statistic":18,"alignment":88},"final translation error ||u||","m","first-pose",{"label":90,"unit":39,"statistic":91,"alignment":88},"DM translation","mean",{"label":93,"unit":94,"statistic":18,"alignment":88},"final rotation error ||omega||","rad",{"label":96,"unit":39,"statistic":91,"alignment":88},"DM rotation",{"label":98,"unit":39,"statistic":91,"alignment":88},"DM (full pose)",[100,104,107,110,113,115,118],{"dataset":101,"sequence":102,"environment":103},"Challenging data sets (ETH)","Apartment (22 m)","indoor apartment",{"dataset":101,"sequence":105,"environment":106},"Hauptgebaude (24 m)","indoor, structured (Hauptgebaude)",{"dataset":101,"sequence":108,"environment":109},"Stairs (12 m)","indoor, structured (Stairs)",{"dataset":101,"sequence":111,"environment":112},"Gazebo Summer (14 m)","semi-structured outdoor (Gazebo)",{"dataset":101,"sequence":114,"environment":112},"Gazebo Winter (15 m)",{"dataset":101,"sequence":116,"environment":117},"Wood Autumn (18 m)","unstructured forest",{"dataset":101,"sequence":119,"environment":117},"Wood Summer (21 m)",[121,124],{"name":122,"methodId":5,"linkable":123,"proposed":123,"self":123},"CELLO-3D (ICP odometry)",true,{"name":7,"methodId":5,"linkable":123,"proposed":123,"self":123},[126,130,133,136,138,141,143,145,147,149,151,153,155,157,159,161,163,165,167,168,170,172,173,175,177,179,182,184,186,188,190,193,195,197,199],[127,127,127,128,129,127,129,129,127],0,0.115,-1,[131,131,127,132,129,127,129,129,131],1,0.274,[127,134,127,135,129,127,129,129,131],2,0.0331,[131,77,127,137,129,127,129,129,131],0.16,[131,139,127,140,129,127,129,129,131],4,0.54,[127,127,131,142,129,127,129,129,131],0.168,[131,131,131,144,129,127,129,129,131],0.467,[127,134,131,146,129,127,129,129,131],0.0091,[131,77,131,148,129,127,129,129,131],0.346,[131,139,131,150,129,127,129,129,131],1.15,[127,127,134,152,129,127,129,129,131],0.0664,[131,131,134,154,129,127,129,129,131],0.0998,[127,134,134,156,129,127,129,129,131],0.0127,[131,77,134,158,129,127,129,129,131],0.307,[131,139,134,160,129,127,129,129,131],0.592,[127,127,77,162,129,127,129,129,131],0.0396,[131,131,77,164,129,127,129,129,131],0.278,[127,134,77,166,129,127,129,129,131],0.0165,[131,77,77,164,129,127,129,129,131],[131,139,77,169,129,127,129,129,131],0.491,[127,127,139,171,129,127,129,129,131],0.0311,[131,131,139,134,129,127,129,129,131],[127,134,139,174,129,127,129,129,131],0.0144,[131,77,139,176,129,127,129,129,131],2.9,[131,139,139,178,129,127,129,129,131],2.5,[127,127,180,181,129,127,129,129,131],5,0.217,[131,131,180,183,129,127,129,129,131],0.205,[127,134,180,185,129,127,129,129,131],0.0178,[131,77,180,187,129,127,129,129,131],0.394,[131,139,180,189,129,127,129,129,131],0.405,[127,127,191,192,129,127,129,129,131],6,0.332,[131,131,191,194,129,127,129,129,131],0.208,[127,134,191,196,129,127,129,129,131],0.0299,[131,77,191,198,129,127,129,129,131],0.533,[131,139,191,200,129,127,129,129,131],0.762,[],[82],[],[],[206,207],"ICP odometry over whole sequence, initial guesses sampled with a = 0.05; final pose error and Mahalanobis distance DM against ground truth; DM averaged over 100 trajectories","Final odometry error and consistency",{"slug":209,"group":210,"sourceId":211,"sourceLabel":212,"table":213,"selfRows":214,"metrics":215,"seqs":221,"entrants":240,"cells":246,"outcomes":315,"locators":317,"hardware":318,"wordings":319,"notes":320},"brossard2020icpcov-table-2","brossard2020icpcov:Table 2","brossard2020icpcov","Brossard et al., 2020","Table 2",16,[216,219],{"label":217,"unit":218,"statistic":91,"alignment":39},"Mah. dist. trans.","unitless",{"label":220,"unit":218,"statistic":91,"alignment":39},"Mah. dist. rot.",[222,226,228,230,232,234,236,238],{"dataset":223,"sequence":224,"environment":225},"Challenging data sets for point cloud registration (Pomerleau et al. 2012)","Apartment","per-sequence type not stated in the paper",{"dataset":223,"sequence":227,"environment":225},"Hauptgebaude",{"dataset":223,"sequence":229,"environment":225},"Stairs",{"dataset":223,"sequence":231,"environment":225},"Mountain",{"dataset":223,"sequence":233,"environment":225},"Gazebo summer",{"dataset":223,"sequence":235,"environment":225},"Gazebo winter",{"dataset":223,"sequence":237,"environment":225},"Wood summer",{"dataset":223,"sequence":239,"environment":225},"Wood winter",[241,242,244],{"name":7,"methodId":5,"linkable":123,"proposed":63,"self":123},{"name":243,"methodId":42,"linkable":63,"proposed":63,"self":63},"ini.+ICP (fusion without cross-covariance)",{"name":245,"methodId":211,"linkable":123,"proposed":123,"self":63},"proposed (full ML covariance, Eq. 15)",[247,249,251,253,254,255,256,257,258,259,260,261,262,263,264,266,267,269,271,273,275,277,279,281,283,284,286,287,289,290,292,293,295,297,299,301,302,303,304,305,306,307,308,309,310,311,313,314],[127,127,127,248,129,127,129,129,127],0.2,[127,131,127,250,129,127,129,129,127],0.1,[127,127,131,252,129,127,129,129,127],0.3,[127,131,131,248,129,127,129,129,127],[127,127,134,250,129,127,129,129,127],[127,131,134,248,129,127,129,129,127],[127,127,77,42,127,127,129,129,127],[127,131,77,42,127,127,129,129,127],[127,127,139,248,129,127,129,129,127],[127,131,139,248,129,127,129,129,127],[127,127,180,250,129,127,129,129,127],[127,131,180,248,129,127,129,129,127],[127,127,191,250,129,127,129,129,127],[127,131,191,252,129,127,129,129,127],[127,127,265,250,129,127,129,129,127],7,[127,131,265,252,129,127,129,129,127],[131,127,127,268,129,127,129,129,127],3.5,[131,131,127,270,129,127,129,129,127],15,[131,127,131,272,129,127,129,129,127],1.9,[131,131,131,274,129,127,129,129,127],3.2,[131,127,134,276,129,127,129,129,127],1.1,[131,131,134,278,129,127,129,129,127],4.2,[131,127,77,280,129,127,129,129,127],1.5,[131,131,77,282,129,127,129,129,127],1.2,[131,127,139,276,129,127,129,129,127],[131,131,139,285,129,127,129,129,127],2.1,[131,127,180,272,129,127,129,129,127],[131,131,180,288,129,127,129,129,127],3.7,[131,127,191,280,129,127,129,129,127],[131,131,191,291,129,127,129,129,127],4.6,[131,127,265,282,129,127,129,129,127],[131,131,265,294,129,127,129,129,127],4.8,[134,127,127,296,129,127,129,129,127],2.3,[134,131,127,298,129,127,129,129,127],9.8,[134,127,131,300,129,127,129,129,127],1.8,[134,131,131,176,129,127,129,129,127],[134,127,134,276,129,127,129,129,127],[134,131,134,278,129,127,129,129,127],[134,127,77,282,129,127,129,129,127],[134,131,77,282,129,127,129,129,127],[134,127,139,131,129,127,129,129,127],[134,131,139,296,129,127,129,129,127],[134,127,180,300,129,127,129,129,127],[134,131,180,288,129,127,129,129,127],[134,127,191,280,129,127,129,129,127],[134,131,191,312,129,127,129,129,127],4.7,[134,127,265,282,129,127,129,129,127],[134,131,265,278,129,127,129,129,127],[316],"not_run (Mountain not considered in the CELLO-3D paper)",[213],[],[],[321],"Trajectory consistency: Mahalanobis distance of compounded ICP trajectories to ground truth, averaged over 40 initial trajectories per sequence; target 1, below 1 pessimistic; CELLO-3D reproduced from Landry et al. with a slightly different ICP setting",{"slug":323,"group":324,"sourceId":5,"sourceLabel":6,"table":325,"selfRows":265,"metrics":326,"seqs":329,"entrants":344,"cells":352,"outcomes":395,"locators":396,"hardware":397,"wordings":398,"notes":399},"landry2019cello3d-table-ii","landry2019cello3d:Table II","Table II",[327],{"label":328,"unit":39,"statistic":91,"alignment":39},"Avg. KL divergence",[330,332,334,336,338,340,342],{"dataset":101,"sequence":331,"environment":103},"Apartment (trained on Haupt. and Stairs; 1190 pairs)",{"dataset":101,"sequence":333,"environment":106},"Hauptgebaude (trained on Apt and Stairs; 938 pairs)",{"dataset":101,"sequence":335,"environment":109},"Stairs (trained on Apt and Haupt.; 798 pairs)",{"dataset":101,"sequence":337,"environment":112},"Gazebo Summer (trained on Gzb. Winter; 826 pairs)",{"dataset":101,"sequence":339,"environment":112},"Gazebo Winter (trained on Gzb. Summer; 798 pairs)",{"dataset":101,"sequence":341,"environment":117},"Wood Autumn (trained on Wd Summer; 812 pairs)",{"dataset":101,"sequence":343,"environment":117},"Wood Summer (trained on Wd Autumn; 966 pairs)",[345,347,349],{"name":346,"methodId":42,"linkable":63,"proposed":63,"self":63},"Baseline (mean training covariance)",{"name":348,"methodId":5,"linkable":123,"proposed":123,"self":123},"Ours (CELLO-3D)",{"name":350,"methodId":351,"linkable":123,"proposed":63,"self":63},"Censi","censi2007covariance",[353,355,357,359,361,363,365,367,369,371,373,375,377,379,381,383,385,387,389,391,393],[127,127,127,354,129,127,129,129,127],34.1,[131,127,127,356,129,127,129,129,131],26.6,[134,127,127,358,129,127,129,129,131],91900000,[127,127,131,360,129,127,129,129,131],34,[131,127,131,362,129,127,129,129,131],26.7,[134,127,131,364,129,127,129,129,131],206000000,[127,127,134,366,129,127,129,129,131],33.7,[131,127,134,368,129,127,129,129,131],27,[134,127,134,370,129,127,129,129,131],76500000,[127,127,77,372,129,127,129,129,131],20.8,[131,127,77,374,129,127,129,129,131],19.8,[134,127,77,376,129,127,129,129,131],2550000,[127,127,139,378,129,127,129,129,131],20.3,[131,127,139,380,129,127,129,129,131],18.9,[134,127,139,382,129,127,129,129,131],2250000,[127,127,180,384,129,127,129,129,131],13.2,[131,127,180,386,129,127,129,129,131],11.5,[134,127,180,388,129,127,129,129,131],4940000,[127,127,191,390,129,127,129,129,131],13.6,[131,127,191,392,129,127,129,129,131],11.3,[134,127,191,394,129,127,129,129,131],35200000,[],[325],[],[],[400,401],"Average KL divergence between sampled and predicted ICP covariance per test group; trained on the named other group","Average KL divergence per test group",[],1790510664479]