[{"data":1,"prerenderedAt":116},["ShallowReactive",2],{"method-gsicpslam2024":3},{"method":4,"reference":62,"equipment":83,"figures":94,"results":88},{"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":33,"platform":35,"estimator":38,"association":39,"timeModel":40,"deskew":41,"loopClosure":42,"globalOptimization":43,"mapRepresentation":44,"prior":45,"outputGeometry":46,"compute":47,"codeUrl":48,"codeLicense":49,"relatedVersions":50},"gsicpslam2024","Ha et al., 2024","GS-ICP SLAM","RGBD GS-ICP SLAM",2024,"recent","C09","odometry_with_local_mapping","GS-ICP SLAM 讓追蹤與建圖共用同一張三維高斯地圖：追蹤端把目前深度影像降採樣反投影後，以 k 近鄰共變異數組成來源高斯，再用廣義 ICP（G-ICP）與地圖中的目標高斯配準求得位姿；建圖端則直接沿用這些共變異數作為新增高斯的初始形狀，並依深度做尺度正規化，因此不需 3DGS 的密化步驟。系統另把追蹤關鍵影格與僅供建圖的關鍵影格分開，以兼顧軌跡精度與渲染品質；整體速度最高達每秒 107 影格，但沒有迴圈閉合，也未評估三維幾何精度。","RGB-D SLAM that tracks with G-ICP against the covariances of a single shared 3D Gaussian map and seeds new Gaussians with the G-ICP covariances (depth-normalized scale), reaching up to 107 FPS; no loop closure and no geometric accuracy evaluation.","full_text_reviewed","peer_reviewed_published","main_body","論文正文只在 Replica 與 TUM RGB-D 驗證，未涉及營建場域，地圖品質僅以渲染指標評估；ECCV 補充資料另報告渲染深度的平均 L1 誤差為 Replica 0.030 m、TUM 0.118 m，仍不是對獨立參考的網格或點雲精度。其追蹤核心是點雲配準中的 G-ICP [segal2009gicp]，實作建立在 VGICP [koide2021vgicp] 上，與雷射 SLAM 的掃描配準一脈相承；語料中的地下工程與室內數位孿生研究以它為 3DGS 基準，包括 yan2026_underground3dgsslam 的地下 RGB-D 實測資料（Table III）與 TUM（Table VI），以及 yuan2026_adaptive3dgsslam（Table 3）。作者也承認深度雜訊會限制真實場景的地圖品質。",[20,21],"public_benchmark","simulation",[23,24,25,26],"Replica average ATE RMSE 0.16 cm versus 0.36 cm for SplaTAM and 0.50 cm for GS-SLAM (Table 1)","Whole-system speed up to 107 FPS on Replica (average 98.11 FPS) and 73.92 FPS on TUM without a tracking limit (Tables 3-4)","TUM average ATE 2.4 cm, the lowest among coupled single-map methods in arXiv v2 Table 2; the ECCV version adds Gaussian Splatting SLAM (MonoGS, keyframes only) at 1.5 cm, so GS-ICP is described there as competitive (ECCV Table 1)","Reusing G-ICP covariances for new Gaussians cuts Replica ATE from 8.893 to 0.157 cm in the ablation (Table 6)",[28,29,30,31,32],"Relies solely on depth for 3D structure, so map quality in real environments is limited by RGB-D depth noise (Sec. 5)","On TUM, PSNR is about 11.7% lower than SplaTAM, and decoupled ORB-SLAM3 and Photo-SLAM track better (1.3 cm versus 2.4 cm) (Sec. 4.2-4.3, Tables 2 and 4)","Without bundle adjustment its rendered-depth quality is lower than GS-SLAM (Replica depth L1 0.030 m versus 0.012 m reported by GS-SLAM) (ECCV Supp. Sec. C.2)","Adding many tracking keyframes degrades accuracy through accumulated scan-matching error (Sec. 3.2, Table 7)","No loop closure, and the main text assesses map quality by rendering metrics only (Sec. 3-4) (inference)",[34],"RGB-D camera",[36,37],"simulation (Replica synthetic scenes)","not_reported (TUM RGB-D capture platform not described)","Generalized-ICP (G-ICP) scan-to-map registration: maximum-likelihood alignment of source Gaussians from the downsampled, reprojected depth image to target Gaussians taken from the 3DGS map, with ellipse scale regularization; 3DGS mapping runs in parallel with L1 and D-SSIM colour and L1 depth losses (Sec. 3, Eq. 1-3); the G-ICP module is built on the VGICP C++ implementation [koide2021vgicp] wrapped with pybind11 (ECCV Supp. Sec. A)","nearest-neighbour correspondences between source and target Gaussians inside G-ICP; keyframes selected when the share of correspondences within a distance threshold falls below a threshold, plus mapping-only keyframes every 10 frames (Sec. 3.1-3.2)","discrete poses","not_applicable","none","none; mapping trains on one randomly chosen past keyframe per iteration and prunes degenerate Gaussians, but poses are not re-optimized globally (Sec. 3.2)","single 3D Gaussian map shared by tracking and mapping: new Gaussians inherit G-ICP k-nearest-neighbour covariances with depth-dependent scale normalization (divided by z^p, p = 1.5 best), without densification (Sec. 3.2, Tables 6 and 8)","none; geometry comes from the sensor depth only (Sec. 5)","3D Gaussian map rendered to colour and depth; the main text evaluates trajectory (ATE) and rendering (PSNR, SSIM, LPIPS) only (Sec. 4.1); the ECCV supplementary adds average rendered-depth L1 error of 0.030 m on Replica and 0.118 m on TUM, with no mesh or point-cloud accuracy against a reference (Supp. Sec. C.2)","AMD Ryzen 7 7800X3D, 32 GB RAM, NVIDIA RTX 4090 24 GB; whole system up to 107 FPS (average 98.11 FPS on Replica, 73.92 FPS on TUM) without tracking limit, or capped at 30 FPS (Sec. 4.1, Tables 3-4)","https:\u002F\u002Fgithub.com\u002FLab-of-AI-and-Robotics\u002FGS_ICP_SLAM","MIT (LICENSE file checked)",[51,55,58],{"relation":52,"title":53,"doi_or_url":54},"preprint","arXiv:2403.12550v2","https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.12550",{"relation":56,"title":57,"doi_or_url":48},"code_release","Lab-of-AI-and-Robotics\u002FGS_ICP_SLAM",{"relation":59,"title":60,"doi_or_url":61},"supplementary_material","ECCV 2024 electronic supplementary material (implementation details, speed sweep, geometric quality)","https:\u002F\u002Fmedia.springernature.com\u002Foriginal\u002Fspringer-static\u002Fesm\u002Fchp%3A10.1007%2F978-3-031-72764-1_11\u002FMediaObjects\u002F635169_1_En_11_MOESM1_ESM.pdf",{"id":5,"kind":63,"shortName":7,"title":8,"authors":64,"year":9,"venue":68,"venueType":69,"publisher":70,"volumeIssuePages":71,"doi":72,"arxivId":73,"url":74,"firstPublicDate":75,"publicationStatus":16,"metadataStatus":76,"fulltextStatus":15,"era":10,"classicReason":41,"codeUrl":48,"cluster":11,"topics":77,"mdpi":78,"verification":79,"label":6,"fulltextRoute":80,"versionRead":81,"addedByCensus":82},"method",[65,66,67],"Seongbo Ha","Jiung Yeon","Hyeonwoo Yu","Computer Vision - ECCV 2024 (Lecture Notes in Computer Science)","conference","Springer Nature Switzerland","pp. 180-197","10.1007\u002F978-3-031-72764-1_11","2403.12550","https:\u002F\u002Fapi.crossref.org\u002Fworks\u002F10.1007\u002F978-3-031-72764-1_11","2024-03-19","metadata_verified",[11],false,"corrected","arXiv","arXiv v2 (2403.12550v2, 22 Mar 2024, CC BY 4.0) read in full; the ECCV 2024 LNCS version of record and its electronic supplementary material (Springer, NTU institutional access) were also read and compared",true,[84,91],{"category":85,"model":86,"canonical":86,"role":87,"dataset":88,"specs":89,"locator":90},"compute","Ryzen 7 7800x3d","compute for runtime",null,"CPU, desktop with 32GB RAM","Sec. 4.1",{"category":85,"model":92,"canonical":92,"role":87,"dataset":88,"specs":93,"locator":90},"NVIDIA RTX 4090 24GB","24 GB GPU",[95,108],{"refId":5,"refLabel":6,"fig":96,"whatZh":97,"license":98,"licenseUrl":99,"sourceUrl":100,"src":101,"width":102,"height":103,"thumb":104,"thumbWidth":105,"thumbHeight":106,"modified":107},"Fig. 1","各稠密表示 SLAM 的影像品質（PSNR）與整體系統 FPS 比較圖","CC BY 4.0 (arXiv v2 version of the figure; the ECCV version is published by Springer under exclusive licence)","https:\u002F\u002Fcreativecommons.org\u002Flicenses\u002Fby\u002F4.0\u002F","https:\u002F\u002Farxiv.org\u002Fhtml\u002F2403.12550v2\u002Ftitle.png","\u002Ffigure-files\u002Fgsicpslam2024\u002Ffig-1.webp",1400,802,"\u002Ffigure-files\u002Fgsicpslam2024\u002Ffig-1.thumb.webp",480,275,"resized to at most 1400 px wide and converted to WebP",{"refId":5,"refLabel":6,"fig":109,"whatZh":110,"license":98,"licenseUrl":99,"sourceUrl":111,"src":112,"width":102,"height":113,"thumb":114,"thumbWidth":105,"thumbHeight":115,"modified":107},"Fig. 4","追蹤關鍵影格與僅供建圖關鍵影格分開選取的三種情境示意及其 ATE 與 PSNR","https:\u002F\u002Farxiv.org\u002Fhtml\u002F2403.12550v2\u002Ffig1.png","\u002Ffigure-files\u002Fgsicpslam2024\u002Ffig-4.webp",437,"\u002Ffigure-files\u002Fgsicpslam2024\u002Ffig-4.thumb.webp",150,1790510663957]