NICER-SLAM
NICER-SLAM 是只用單眼 RGB 影像的神經隱式 SLAM，追蹤與建圖共用同一個階層式 SDF 表示：粗層為 32 立方的稠密特徵格網，細層以多解析度網格學習殘差 SDF，另以多解析度網格表示顏色。因為沒有深度量測，建圖時額外加入 Omnidata 的單眼深度與法向量、GMFlow 光流、影像扭曲與 Eikonal 等損失來消除歧義，並依各體素取樣次數局部調整 SDF 轉密度的參數。系統在 Replica 上的幾何品質接近 RGB-D 方法，但追蹤不如 DROID-SLAM，未做迴圈閉合，也遠非即時。
本頁內容
RGB-only neural implicit SLAM that tracks and maps with one hierarchical SDF and colour grid, disambiguated by monocular depth and normal priors, optical flow and a warping loss, with a locally adaptive SDF-to-density transform; no loop closure and not real-time.
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 |
|---|---|
| 原文測試平台 | simulation (Replica)、原文未報告 (7-Scenes and the self-captured outdoor Azure Kinect dataset; carrier not described) |
| 狀態估計 | end-to-end optimization through differentiable volume rendering: tracking optimizes the current pose with an RGB rendering loss (100 iterations, 1024 pixels) with the map fixed; mapping runs a 3-stage optimization with RGB, warping, optical-flow, monocular depth, monocular normal and Eikonal losses, ending with local bundle adjustment over 16 selected frames of which half are frozen (Sec. 3.3; iteration counts from arXiv v1 Sec. 3.4) |
| 資料關聯 | direct photometric rendering loss plus dense correspondence cues: RGB warping between keyframes and optical flow from GMFlow (Sec. 3.3) |
| 時間表示 | discrete poses |
| 去畸變 | 不適用 |
| 迴圈閉合 | none (stated as a limitation, Sec. 5) |
| 全域最佳化 | none; local BA over selected mapping frames only (Sec. 3.3-3.4) |
| 地圖表示 | hierarchical neural implicit SDF: coarse 32^3 dense feature grid plus 8-level fine residual grids (32-128) and a 16-level colour grid (16-2048) with small MLP decoders; VolSDF-style SDF-to-density with a locally adaptive beta from per-voxel sample counts (Sec. 3.1-3.2) |
| 先驗資訊 | monocular depth and normal predictions from an off-the-shelf predictor (Omnidata in arXiv v1), optical flow from GMFlow; COLMAP used to obtain intrinsics for 7-Scenes and the self-captured outdoor dataset (Sec. 3.3, Sec. 4) |
| 可輸出幾何 | camera trajectory and a triangle mesh extracted by marching cubes at 512^3; rendered novel views (Sec. 3.4) |
| 計算需求 | not reported in the 3DV main text; arXiv v1 Sec. 3.4 reports a single NVIDIA A100 with on average 496 ms per mapping iteration and 147 ms per tracking iteration (100 iterations each); not real-time (Sec. 5) |
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 |
|---|---|---|---|---|---|
| RGB-D 相機 | Azure Kinect | 方法輸入 | SCO (self-captured outdoor) | used to capture the self-captured outdoor (SCO) dataset of 6 scenes with 800 to 2700 frames; only RGB images are input, the depth is shown for visualization and is unreliable outdoors | (Zhu et al., 2024, Sec. 4 Datasets; Sec. 4.1; Fig. 6) |
| 運算硬體 | A100 | 執行運算平台 | 未標示 | single GPU; 496 ms per mapping iteration, 147 ms per tracking iteration | (Zhu et al., 2024, arXiv v1 Sec. 3.4 (implementation details are not in the 3DV main text)) |
作者報告的優勢與限制
優勢
- Best Replica geometry among RGB-only methods (accuracy 3.65 cm, completion 4.16 cm, completion ratio 79.37%, normal consistency 90.27%), ahead of TANDEM, NeRF-SLAM, DIM-SLAM* and DROID-SLAM and close to RGB-D NICE-SLAM (3.87, 3.87, 82.41, 89.93) (Table 1)
- Replica tracking on par with RGB-D NICE-SLAM (1.88 versus 1.95 cm average ATE) without depth input (Table 3, Sec. 4.1)
- Best average novel-view PSNR on Replica for both extrapolated (23.93 dB) and interpolated (25.41 dB) views, above RGB-D NICE-SLAM (23.26 and 24.42 dB) (Table 2)
- On the self-captured outdoor Azure Kinect scenes it reconstructs textureless walls and small details such as a handrail where most monocular baselines fail (qualitative, Sec. 4.1, Fig. 6)
限制
- Not optimized for real-time operation (Sec. 5); about 496 ms per mapping iteration on an A100 (arXiv v1 Sec. 3.4)
- No loop closure, so tracking could be improved (Sec. 5)
- Tracking is clearly worse than systems built for tracking: Replica 1.88 cm versus 0.33 (DROID-SLAM), 0.55 (DSO), 0.46 (DIM-SLAM) and 1.15 cm (TANDEM) (Table 3); 7-Scenes 8.55 versus 5.66 cm for DROID-SLAM (arXiv v1 Table 4)
- Removing the monocular depth or normal loss sharply degrades mapping and tracking, showing dependence on learned priors (arXiv v1 Table 5a; the 3DV version moves ablations to the supplementary)
- The outdoor evaluation is qualitative only, and the Azure Kinect depth is unreliable outdoors (Sec. 4.1, Fig. 6)
營建工程相關證據
論文未涉及營建場域；定量評估只在合成 Replica 資料，3DV 版另以 Azure Kinect 自行拍攝 6 個戶外場景，但只做定性比較，且作者指出戶外深度量測不可靠。其以單眼深度與法向量先驗補足 RGB 幾何的做法，說明僅靠影像取得公分級網格仍依賴學習先驗，且未做迴圈閉合、速度遠非即時；對以一般相機記錄工地的情境可作為上限參考，但無法支持施工驗收等級的幾何主張（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
Zhu et al., 2024
（2024）NICER-SLAM: Neural Implicit Scene Encoding for RGB SLAM2024 International Conference on 3D Vision (3DV), pp. 42-52
DOI 10.1109/3dv62453.2024.00096arXiv 2302.03594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
- version of record：3DV 2024 proceedings (IEEE Xplore 10550721) https://doi.org/10.1109/3DV62453.2024.00096
- 預印本：arXiv:2302.03594v1 https://arxiv.org/abs/2302.03594
- 程式碼釋出：cvg/nicer-slam https://github.com/cvg/nicer-slam
程式碼：https://github.com/cvg/nicer-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。