SLAM3R
SLAM3R 以前饋式神經網路直接從單眼 RGB 影片產生稠密點雲，而不求解任何相機參數。影片先以滑動視窗切成重疊片段，影像對點雲（I2P）網路以多視角交叉注意力，從 11 張影像回歸視窗中間關鍵影格的點雲；局部對世界（L2W）網路再參考以檢索模組從儲存池挑出的歷史影格，把局部點雲逐步配準到全域座標。兩個網路都以 DUSt3R 權重初始化，在單張 4090D 上約每秒 24 至 25 影格，但因沒有相機參數，無法做全域光束法平差，推導出的位姿也不如專門的 SLAM。
本頁內容
Feed-forward monocular dense reconstruction that regresses keyframe pointmaps from sliding-window clips (I2P) and registers them into a global frame with retrieved scene frames (L2W), all initialized from DUSt3R and running at over 20 FPS without explicit camera parameters or bundle adjustment.
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 RGB camera (video) |
|---|---|
| 原文測試平台 | 原文未報告 (7-Scenes, ScanNet, ETH3D and in-the-wild videos; capture platforms not described)、simulation (Replica) |
| 狀態估計 | feed-forward networks without explicit camera parameters: an Image-to-Points (I2P) ViT with multi-view cross-attention regresses keyframe pointmaps from sliding-window clips (L = 11), and a Local-to-World (L2W) network registers each keyframe's pointmap into the global frame using retrieved scene frames; no pose optimization or bundle adjustment (Sec. 3) |
| 資料關聯 | implicit through cross-attention between keyframe and supporting or scene-frame tokens; a learned retrieval module selects the top-K scene frames from a reservoir of registered frames (Sec. 3.2, Supp. A) |
| 時間表示 | 不適用 (no explicit camera poses; poses can be derived afterwards by PnP-RANSAC) |
| 去畸變 | 不適用 |
| 迴圈閉合 | none as optimization; retrieval of long-term scene frames acts as implicit re-localization during registration (Sec. 4.2) |
| 全域最佳化 | none; authors state that removing camera parameters prevents global bundle adjustment (Sec. 5) |
| 地圖表示 | global dense point cloud built from registered pointmaps with per-pixel confidence (confidence threshold 3 in evaluation) (Sec. 3, Supp. B) |
| 先驗資訊 | networks initialized from DUSt3R (224 x 224) weights and trained on about 850K clips from ScanNet++, Aria Synthetic Environments and CO3D-v2 (Sec. 4, Supp. A) |
| 可輸出幾何 | dense 3D point cloud (colored by input frames); camera poses only as a derived by-product via PnP-RANSAC (Sec. 4.1) |
| 計算需求 | about 24-25 FPS on a single NVIDIA 4090D at 224 x 224 input; training on 8 NVIDIA 4090D (24 GB) for about one day (Sec. 4, Tables 1-2) |
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 | NVIDIA 4090D | 執行運算平台 | 未標示 | single GPU for FPS measurements; training on 8 GPUs with 24 GB each | (Liu et al., 2025, Sec. 4, Sec. 4.1) |
作者報告的優勢與限制
優勢
- Best accuracy and completeness among real-time methods on 7 Scenes (average 2.13 / 2.34 cm at about 25 FPS) (Table 1)
- On Replica, accuracy and completeness (3.57 / 2.62 cm) comparable to optimization-based NICER-SLAM and DUSt3R while running at about 24 FPS (Table 2)
- Learned L2W registration with retrieval is more accurate and faster than DUSt3R global alignment or Umeyama plus ICP (Table 5)
- Much less drift than the concurrent Spann3R (Replica ATE 6.61 versus 32.79 cm) (Table 3, Supp. Table 7)
限制
- No camera parameters, so global bundle adjustment cannot be performed (Sec. 5)
- Derived poses fall short of dedicated SLAM systems (Replica ATE 6.61 cm versus 0.33 cm for DROID-SLAM) (Sec. 5, Table 3)
- Input images are center-cropped to 224 x 224 (Sec. 4)
- Evaluation aligns reconstructions to ground truth with Umeyama and ICP, so metric scale and absolute drift are not assessed directly (Sec. 4.1) (inference)
營建工程相關證據
論文未涉及營建場域；定量評估限於 7-Scenes、Replica 與補充材料中 ScanNet、Tanks and Temples、ETH3D 的少數場景，且重建先以 Umeyama 與 ICP 對齊真值再計算公分級誤差，並未檢驗公制尺度或以獨立量測驗證。其不需相機參數即可即時由影片產生稠密點雲，對以一般手機或相機快速記錄工地狀況有吸引力，但輸入解析度低、無全域平差，量測用途仍需另行校核尺度與精度（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
Liu et al., 2025
（2025）SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 16651-16662
DOI 10.1109/cvpr52734.2025.01552arXiv 2412.09401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:2412.09401v3 https://arxiv.org/abs/2412.09401
- 程式碼釋出：PKU-VCL-3DV/SLAM3R https://github.com/PKU-VCL-3DV/SLAM3R
程式碼：https://github.com/PKU-VCL-3DV/SLAM3R（授權：CC BY-NC-SA 4.0 per the repository LICENSE text (GitHub API reports NOASSERTION)）。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。