GO-SLAM
GO-SLAM 以 DROID-SLAM 的學習式稠密光流與可微分稠密光束法平差（dense bundle adjustment）作為追蹤核心，在前端依光流估算的共視度偵測迴圈，並在獨立執行緒中對所有關鍵影格線上執行完整光束法平差，以抑制長序列的累積漂移。建圖端使用多解析度雜湊編碼的神經隱式 SDF，每次優先選取位姿變化最大的關鍵影格重新訓練，使重建隨全域最佳化後的位姿與深度同步更新。同一架構可接受單眼、立體或 RGB-D 影像，最後以 marching cubes 從 SDF 擷取網格。
本頁內容
Extends DROID-SLAM with flow-based loop closing and online full bundle adjustment, and continuously re-fits a hash-encoded neural SDF to the globally optimized keyframe poses and depths, giving globally consistent meshes from monocular, stereo or RGB-D video.
技術屬性
欄位內容為文獻擷取紀錄的原文用語（英文），以原文為據；「未查證」表示本研究尚未讀到該資訊，不代表該方法不具備此能力。
| 感測輸入 | monocular camera、stereo camera、RGB-D camera |
|---|---|
| 原文測試平台 | UAV (EuRoC MAV dataset)、simulation (Replica synthetic scenes)、原文未報告 (capture platforms of TUM RGB-D, ETH3D-SLAM and ScanNet are not described in the paper) |
| 狀態估計 | DROID-SLAM tracking extended with online global optimization: RAFT-based recurrent update operator predicts dense flow and confidence, and a differentiable dense bundle adjustment (DBA) layer solves poses and per-pixel inverse depths by damped Gauss-Newton; front end optimizes a local keyframe window with loop edges, back end runs online full BA over all keyframes in a separate thread (Sec. 3.1) |
| 資料關聯 | dense learned optical flow with per-pixel confidence between keyframe pairs; keyframe-graph edges chosen by co-visibility measured as mean rigid flow (threshold tau_co = 25) with neighbourhood suppression (Sec. 3.1, Sec. 4.1) |
| 時間表示 | discrete poses (keyframes) |
| 去畸變 | 不適用 |
| 迴圈閉合 | flow-based: edges sampled from the unexplored part of the co-visibility matrix between local-window and historical keyframes; a loop is accepted after three consecutive candidates with mean flow below tau_co, then optimized by DBA (Sec. 3.1) |
| 全域最佳化 | online full bundle adjustment over all keyframes in a back-end thread, running concurrently with tracking and loop closing (Sec. 3.1) |
| 地圖表示 | neural implicit SDF and colour field with multi-resolution hash encoding (16 levels, Instant-NGP style) and shallow MLPs, rendered by NeuS-style unbiased volume rendering; mapping re-trains on selected keyframes (latest two, top-10 by pose change, 10 stratified) (Sec. 3.2, Supp. A) |
| 先驗資訊 | pretrained DROID-SLAM weights for tracking; rendering networks trained from scratch per scene (Sec. 4.1) |
| 可輸出幾何 | keyframe trajectory, per-keyframe depth and a triangle mesh extracted by marching cubes from the SDF (Sec. 4.1) |
| 計算需求 | Intel Core i9-10920X 3.5 GHz CPU and NVIDIA RTX 3090 GPU; about 8 FPS on Replica RGB-D with 15.63 GB GPU memory in Table 9, while Sec. 4.4 states a maximum of 18 GB (Sec. 4.1, Sec. 4.4, Table 9) |
使用設備
原文使用的感測器、運算硬體與載具（equipment）。型號保留原文寫法，連結到設備頁中同一型號的歸併名稱；角色依原文用途分為方法輸入、資料集感測器、執行運算平台、參考或真值量測（reference or ground truth）與比較對象設備。
| 類別 | 型號（原文寫法） | 角色 | 資料集 | 原文規格 | 出處 |
|---|---|---|---|---|---|
| 運算硬體 | Intel Core i9-10920X | 執行運算平台 | 未標示 | 3.5 GHz CPU | (Zhang et al., 2023b, Sec. 4.1) |
| 運算硬體 | NVIDIA RTX 3090 | 執行運算平台 | 未標示 | GPU; 15.63 GB peak memory used on Replica RGB-D (Table 9) | (Zhang et al., 2023b, Sec. 4.1; Table 9) |
作者報告的優勢與限制
優勢
- Online loop closing and full BA reduce average ATE on 8 ScanNet RGB-D scenes from 11.59 cm (no LC, no full BA) to 7.02 cm, with speed falling from 30 to 10 FPS (Table 8)
- On long monocular ScanNet sequences the average ATE is 17.59 cm against 52.60 cm for DROID-SLAM with final global BA (Table 3)
- One architecture handles monocular, stereo and RGB-D input; stereo EuRoC average ATE 0.024 m equals DROID-SLAM (Table 2)
- Mapping quality degrades little when skipping up to 7 of 8 frames (Replica F-score 85.56 to 84.41; ScanNet ATE 7.02 to 7.28 cm) (Table 7)
限制
- No limitations section; runtime is about 8 FPS on Replica with an RTX 3090 and 15.63 GB (Table 9) or up to 18 GB (Sec. 4.4) of GPU memory, so real-time use relies on frame skipping (Table 7) (inference)
- Monocular accuracy remains scene dependent: ScanNet scene0465 still has 79.51 cm ATE (Table 3)
- Reconstruction metrics are reported only on synthetic Replica; real-scene geometry is shown qualitatively (Sec. 4.4, Figs. 5, B-D) (inference)
- Tracking depends on pretrained DROID-SLAM weights; generalisation beyond indoor and MAV datasets is not tested (Sec. 4.1) (inference)
營建工程相關證據
論文未涉及營建場域；驗證資料為 TUM RGB-D、EuRoC、ETH3D-SLAM、ScanNet 與 Replica 等室內或飛行器資料集，幾何精度只在 Replica 合成場景以公分級 Accuracy 與 Completion 評估，未使用全測站或 TLS 等獨立參考量測。其長序列迴圈閉合與全域平差可說明神經隱式地圖要維持全域一致須仰賴傳統 SLAM 式的後端，但單眼 ScanNet 部分場景仍有數十公分的 ATE（Table 3），距營建量測所需精度仍有差距（推論）。
原文驗證環境：公開基準、模擬
報告的性能數據
性能數據仍在分批查證，目前尚未收錄此方法的報告值。
來源
Zhang et al., 2023b
（2023）GO-SLAM: Global Optimization for Consistent 3D Instant Reconstruction2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 3704-3714
DOI 10.1109/iccv51070.2023.00345arXiv 2309.02436程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 預印本：arXiv:2309.02436v1 https://arxiv.org/abs/2309.02436
- 程式碼釋出：youmi-zym/GO-SLAM https://github.com/youmi-zym/GO-SLAM
程式碼：https://github.com/youmi-zym/GO-SLAM（授權：Apache-2.0 (LICENSE file checked)）。有公開程式碼不等於已被重現，也不代表目前版本與論文版本相同。