PaperlayerView exact revision on arXiv

Source-linked AI summary

How NeRFs and 3D Gaussian Splatting are Reshaping SLAM: a Survey

Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R. Oswald, Matteo Poggi

arXiv:2402.13255v3cs.CVcs.RO

TL;DR

With no comprehensive survey focused on recent radiance-field SLAM, this paper reviews 80 contemporary systems. Its structured synthesis compares developments across tracking, mapping, and rendering and identifies limitations and open challenges.

  • Problem

    The paper identifies a lack of comprehensive surveys examining SLAM methods built on recent radiance-field advances.

    Source: §I · p. 2

  • Method

    The survey investigates 80 systems published in the preceding three years, categorizing methods and tracing their progression.

    Source: §I · p. 2

  • Results

    The synthesis classifies innovations and limitations and presents comparative results across tracking, mapping, and rendering.

    Source: §VI · p. 21

  • Takeaways & Limitations

    The review identifies open challenges as avenues for future exploration.

    Source: §VI · p. 21

    Current 3DGS-style SLAM approaches face scalability and precise-geometry limitations, possible uncontrolled Gaussian growth into unobserved areas, and higher computational complexity than NeRF-based methods.

    Source: §V · p. 20

Abstract

from arXiv · show

Over the past two decades, research in the field of Simultaneous Localization and Mapping (SLAM) has undergone a significant evolution, highlighting its critical role in enabling autonomous exploration of unknown environments. This evolution ranges from hand-crafted methods, through the era of deep learning, to more recent developments focused on Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (3DGS) representations. Recognizing the growing body of research and the absence of a comprehensive survey on the topic, this paper aims to provide the first comprehensive overview of SLAM progress through the lens of the latest advancements in radiance fields. It sheds light on the background, evolutionary path, inherent strengths and limitations, and serves as a fundamental reference to highlight the dynamic progress and specific challenges.

I. INTRODUCTION

SLAM jointly builds maps of unknown environments and estimates a machine’s position and orientation. This survey reviews radiance-field-inspired SLAM amid an identified gap in recent survey coverage.

I. INTRODUCTION

  • SLAM enables machines to navigate unknown environments while incrementally mapping them and estimating their own position and orientation.Source: §I · p. 1
  • SLAM methods progressed from hand-crafted algorithms with real-time performance and scalability to deep-learning methods that improve localization and mapping in challenging conditions.Hand-crafted methods struggle with illumination, radiometric change, dynamic or poorly textured scenes; deep methods depend on extensive training data and accurate ground-truth annotations.Source: §I · pp. 1–2
  • Radiance-field approaches are presented as addressing discrete-map shortcomings with continuous surfaces, improved noise and outlier handling, hole filling, and denser, compact maps that may support arbitrary-resolution meshes.The field remains at an early stage, with each technique having strengths and specific limitations.Source: §I · p. 2
  • The survey reviews 80 radiance-field-inspired SLAM systems published in the preceding three years, categorizing methods, tracing their progression, and examining their requirements.It aims to address the lack of surveys focused on the latest developments.Source: §I · p. 2

A. Existing SLAM Surveys

  • Earlier surveys cover traditional, visual, multi-robot, sensor-specific, dynamic-environment, and deep-learning SLAM, but the paper identifies no comprehensive survey of methods rooted in recent radiance-field advances.The cited survey landscape also includes work on RGB-D reconstruction and active SLAM.Source: A. Existing SLAM Surveys · p. 2
  • Prior radiance-field SLAM work draws on frame-to-frame tracking, often paired with loop closure or global bundle adjustment, and frame-to-model tracking, which can be less scalable to large scenes.The survey relates these strategies to decoupled and coupled methods, respectively.Source: A. Existing SLAM Surveys · pp. 2–3

B. Progress in Radiance Field Theory

Radiance fields represent scene appearance and geometry implicitly, explicitly, or through hybrid combinations. NeRF uses a neural scene function and ray-based volume rendering, with speed and real-time-use constraints that motivate subsequent methods.

B. Progress in Radiance Field Theory

  • Radiance fields may be implicit neural encodings, explicit spatial structures, or hybrids combining stored local features with neural networks.Explicit representations offer faster access but use more memory and have resolution constraints; implicit encodings are compact but can require more rendering computation.Source: B. Progress in Radiance Field Theory · p. 3

1) Neural Radiance Field (NeRF):

  • NeRF represents a scene with an MLP that maps 3D coordinates and viewing direction to color and volume density, using input-view camera poses and intrinsics.The original implementation obtains camera poses and intrinsics using COLMAP.Source: 1) Neural Radiance Field (NeRF) · p. 3
  • NeRF synthesizes an image by casting camera rays, evaluating color and density at sampled points, and integrating their contributions through volume rendering.The integral is approximated using evenly spaced bins along each ray and alpha compositing.Source: 1) Neural Radiance Field (NeRF) · p. 3
  • Some methods use expected depth to add supervision from external priors or regularize the scene, aiming to improve geometry and enforce depth smoothness.Source: 1) Neural Radiance Field (NeRF) · p. 3
  • NeRF variants retain slow training and rendering, while some pose-estimation approaches process image sets together or cannot jointly optimize poses and the network, limiting real-time SLAM use.Structured-grid methods also struggle with empty spaces and image quality limitations that impede rendering speed.Source: 1) Neural Radiance Field (NeRF) · p. 4

2) Surface Reconstruction from Neural Fields:

Neural-field scene geometry can be extracted using occupancy or signed-distance representations rather than relying only on density thresholding. The surveyed alternatives modify how surfaces and their contributions are represented during rendering.

2) Surface Reconstruction from Neural Fields:

  • A basic way to obtain coarse geometry from NeRF is to threshold the density predicted by its MLP.Source: 2) Surface Reconstruction from Neural… · p. 4
  • Occupancy represents free versus occupied space with a learned discrete function, estimates occupancy probabilities, and extracts surfaces using marching cubes.Source: 2) Surface Reconstruction from Neural… · p. 4
  • Signed distance functions encode a point’s distance to the nearest surface, with negative values inside objects and positive values outside; NeuS predicts an SDF and replaces NeRF’s alpha term with an SDF-derived weighting.Truncated SDF rendering can remove contributions from values too far from individual surfaces; one approach obtains pixel color as a weighted sum of ray-sampled colors.Source: 2) Surface Reconstruction from Neural… · p. 4

3) 3D Gaussian Splatting (3DGS):

3D Gaussian Splatting represents scenes with learnable Gaussian primitives and renders them by projecting and blending splats. The section also describes datasets used in SLAM evaluation, spanning indoor RGB-D scenes and autonomous-driving and outdoor settings.

3) 3D Gaussian Splatting (3DGS):

  • 3DGS represents scenes with learnable 3D Gaussian primitives, offering continuous scene modeling and real-time tile-based rendering without neural components or empty-space computation.Each primitive is parameterized by a covariance matrix, position, opacity, and spherical-harmonic color; the approach is contrasted with NeRF’s computationally expensive volumetric ray sampling.Source: 3) 3D Gaussian Splatting (3DGS) · p. 4
  • Rendering projects 3D Gaussians onto the image plane and blends overlapping splats in depth order to produce pixel color and depth.Optimization uses SGD with L1 and D-SSIM losses; adaptive densification adjusts high-gradient points and removes low-opacity points.Source: 3) 3D Gaussian Splatting (3DGS) · pp. 4–5

C. Datasets

  • TUM RGB-D provides 39 color-depth sequences with motion-capture trajectories, while ScanNet supplies 2.5 million images from 1,513 scans across 707 indoor spaces.TUM sequences include handheld and robot motion; ScanNet also provides camera poses, reconstructions, semantic segmentations, and aligned CAD models.Source: C. Datasets · p. 5
  • Replica offers 18 photorealistic indoor scenes, KITTI covers vehicle-based driving data, and Newer College pairs a 2.2 km walk with multimodal sensors and survey-grade LiDAR.Replica’s SLAM sequences are synthetically generated from reconstructed meshes; Newer College provides scan poses accurate to approximately 3 cm.Source: C. Datasets · pp. 5–6

1) Other Datasets:

The section surveys alternative SLAM datasets and summarizes measures for reconstruction, depth, trajectory, view synthesis, and semantic segmentation. The overview table organizes systems by sensor modality and design properties.

1) Other Datasets:

  • ETH3D-SLAM supports visual-inertial mono, stereo, and RGB-D evaluation, while EuRoC MAV provides synchronized stereo, IMU, and accurate ground truth for micro aerial vehicles.ETH3D-SLAM includes 56 training datasets and 35 test datasets; EuRoC includes an industrial setting with millimeter-accurate ground truth.Source: 1) Other Datasets · p. 6
  • Other datasets cover relocalization challenges, high-resolution indoor semantics, synthetic sensor artifacts, and dynamic-scene SLAM with people manipulating objects.These include 7-scenes, ScanNet++, NeuralRGBD, and the Bonn Dataset.Source: 1) Other Datasets · p. 6

D. Evaluation Metrics

  • Reconstruction metrics include Accuracy and Completion distances, Precision and Recall within a distance threshold, and their harmonic-mean F-Score.The section also defines L1-Depth using sampled viewpoints, comparing reconstructed and ground-truth depth maps.Source: D. Evaluation Metrics · p. 6
  • ATE measures average Euclidean translation error between aligned estimated and ground-truth trajectories, while PSNR, SSIM, and LPIPS assess view-synthesis image quality.ATE focuses only on translation; PSNR and SSIM are higher-is-better, while LPIPS is lower-is-better.Source: D. Evaluation Metrics · p. 6
  • Table I organizes SLAM systems by RGB-D, RGB, and LiDAR categories, then records modalities, mapping, outputs, tracking, design strategies, priors, and code links.A dagger marks methods whose code or webpage has not been released.Source: Table I · p. 7
  • Semantic segmentation is evaluated with mIoU, the average of per-class intersection-over-union scores, with higher values indicating better segmentation accuracy.Source: D. Evaluation Metrics · p. 7

III. SIMULTANEOUS LOCALIZATION AND MAPPING

The survey curates peer-reviewed SLAM systems using radiance-field representations and organizes them by sensor type and publication date. It points readers to an overview table and illustrates iMap’s tracking and mapping processes.

III. SIMULTANEOUS LOCALIZATION AND MAPPING

  • The survey groups radiance-field SLAM methods by RGB-D, RGB, and LiDAR sensors, organizing methods chronologically within major conferences and journals.It focuses on peer-reviewed publications as a curated overview and directs readers to recent arXiv submissions for newer developments.Source: §III · p. 7
  • Table I summarizes surveyed methods’ features and provides project-page or source-code links when available.Source: §III · p. 7
  • The iMap illustration contrasts tracking, which optimizes the current frame’s pose in a locked network, with mapping, which jointly refines network parameters and selected-keyframe poses.The mapping illustration uses differentiable rendering functions.Source: Figure 5 · p. 8

A. RGB-D SLAM Approaches

RGB-D radiance-field SLAM includes NeRF-style and 3DGS-style systems, with specialized methods for large scenes, semantics, dynamics, uncertainty, and additional sensors. Across these approaches, designs vary in scene representation, optimization, tracking, and map management.

A. RGB-D SLAM Approaches

  • RGB-D methods divide into NeRF-style and 3DGS-style approaches, with specialized variants for large scenes, semantics, dynamic settings, uncertainty, and event-camera integration.Source: A. RGB-D SLAM Approaches · p. 8

1) NeRF-style RGB-D SLAM:

  • iMAP introduced implicit neural representations to SLAM, jointly optimizing scene-network parameters and selected-keyframe poses while tracking frames in parallel.Its MLP maps 3D coordinates to color and volume density for depth and color rendering.Source: A. RGB-D SLAM Approaches · p. 8
  • NeRF-style RGB-D systems explore hierarchical voxel and feature grids, sparse voxel octrees, multi-scale planes, coordinate and hash encodings, global bundle adjustment, and dynamic neural point clouds.Examples include NICE-SLAM, Vox-Fusion, ESLAM, Co-SLAM, GO-SLAM, and Point-SLAM; GO-SLAM uses parallel tracking and mapping threads and supports monocular, stereo, and RGB-D cameras.Source: 1) NeRF-style RGB-D SLAM · p. 8
  • Further NeRF-style designs address incomplete depth, geometry-aware tracking, planar constraints, generalization to unseen scenes, multiresolution mapping, and low-rank feature storage.Examples include ADFP’s depth-fusion prior, IBD-SLAM’s pretrained depth fusion, HERO-SLAM’s feature-metric optimization, LRSLAM’s tensor decomposition, and LCP-Fusion’s SDF priors.Source: 1) NeRF-style RGB-D SLAM · p. 9

2) 3DGS-style RGB-D SLAM:

  • 3DGS-style RGB-D systems leverage fast photorealistic rendering, expandable Gaussian maps, dense photometric losses, and direct parameter gradients for optimization.Source: 2) 3DGS-style RGB-D SLAM · p. 9
  • MonoGS introduced 3D Gaussians for a moving RGB or RGB-D camera, using direct pose optimization, keyframe management, and pruning of unstable Gaussians.Photo-SLAM combines explicit geometric features with implicit texture representations, while SplaTAM tracks by minimizing re-rendering errors and updates Gaussian parameters.Source: 2) 3DGS-style RGB-D SLAM · p. 9
  • GS-SLAM adaptively adds or removes Gaussians and refines poses coarse-to-fine, while GS-ICP SLAM exchanges Gaussian-map and G-ICP information for tracking and mapping.The bidirectional exchange uses scale alignment to reduce redundant computations.Source: 2) 3DGS-style RGB-D SLAM · pp. 9–10
  • Later 3DGS methods target mapping consistency, uncertainty, inertial inputs, and large-scale real-time operation through regularization, adaptive initialization, or Gaussian opacity roles.HF-GS SLAM guides densification with rendering loss; RTG-SLAM separates opaque surface-fitting Gaussians from transparent residual-color Gaussians.Source: 2) 3DGS-style RGB-D SLAM · p. 10

3) Submaps-based SLAM:

Submap-based systems address dense radiance-field SLAM’s scalability and forgetting challenges through distributed or local maps and coordination strategies. Semantic and dynamic-scene methods extend these representations to scene understanding and settings where static-scene assumptions are problematic.

3) Submaps-based SLAM:

  • Submap-based systems divide large scenes into local or distributed representations, using strategies such as map stitching, distillation, fusion, loop closure, and optimization to manage consistency and scalability.The surveyed approaches include distributed MLPs, neural points, spatial blocks, multiple agents, and incrementally allocated neural submaps; some still face limits on correcting large drift.Source: 3) Submaps-based SLAM · pp. 10–11

4) Semantic RGB-D SLAM:

  • Semantic RGB-D SLAM spans interactive labeling, feature fusion, object-centric models, and semantic neural or Gaussian representations for scene understanding and reconstruction.Examples include user-assisted or uncertainty-guided labels, class-wise scene models, and semantic information embedded in 3D Gaussians.Source: 4) Semantic RGB-D SLAM · pp. 11–12

5) SLAM in Dynamic Environments:

  • Most SLAM methods assume static scenes, while dynamic-scene approaches use motion or semantic cues to identify moving objects and adapt mapping and tracking.The surveyed methods include background inpainting, time-varying representations, motion masks, and dynamic Gaussian management.Source: 5) SLAM in Dynamic Environments · p. 12

6) Uncertainty Estimation:

The section surveys uncertainty-aware SLAM, event-camera systems, and RGB-only methods, including NeRF-style mapping and approaches that use external supervision. These approaches address sensor uncertainty, challenging imaging conditions, and RGB-only depth ambiguity through varied representations and optimization strategies.

6) Uncertainty Estimation:

  • Uncertainty-aware methods model depth noise or sensor uncertainty in mapping and optimization, with approaches using confidence weighting, Bayesian fusion, or visual-inertial integration.The section also notes a need for specialized outdoor datasets with mesh models and well-characterized sensors.Source: 6) Uncertainty Estimation · p. 12

7) Event-based SLAM:

  • Event-camera SLAM is emerging as a response to motion blur and lighting variation; EN-SLAM combines event and RGB-D data through differentiable camera-response rendering.EN-SLAM introduces the DEV-Indoors and DEV-Reals datasets for challenging blur and lighting conditions.Source: 7) Event-based SLAM · pp. 12–13, Figure 9 · p. 13

B. RGB-based SLAM Methodologies

  • RGB-only SLAM applies where depth sensors are impractical, but monocular methods face depth ambiguity and lack geometric priors, which can slow optimization.The section groups RGB-based methods by representation and use of external supervision, semantic estimation, or uncertainty handling.Source: B. RGB-based SLAM Methodologies · p. 13

1) NeRF-style RGB SLAM:

  • NeRF-style RGB SLAM combines implicit scene representations with varied tracking and opacity strategies, including external visual odometry and hybrid rendering.Orbeez-SLAM pairs ORB-SLAM2 pose estimation with Instant-NGP mapping, while TT-HO-SLAM uses ternary opacity and hybrid odometry.Source: 1) NeRF-style RGB SLAM · p. 13
  • MonoGS++ uses DPVO for initial poses and points, then bootstraps and adds 3D Gaussians with planar regularization.Source: 2) 3DGS-style RGB SLAM · p. 13

3) Aided Supervision:

  • Aided-supervision methods incorporate external depth, surface-normal, or optical-flow signals to disambiguate RGB-only optimization and improve SLAM performance.Examples range from depth-supervised multi-threaded mapping and monocular depth priors to joint depth and scale adjustment, geometric representations, and MVS-derived depth for Gaussian mapping.Source: 3) Aided Supervision · pp. 13–14

4) Semantic RGB SLAM:

The surveyed methods extend radiance-field SLAM to semantic understanding, uncertainty handling, and LiDAR-based mapping. The experiments section describes cross-paper evaluation protocols and visual SLAM benchmarks.

4) Semantic RGB SLAM:

  • 3DIML builds 3D-consistent pseudo-labels from image masks before neural label-field training, while RO-MAP performs object-centric monocular mapping without depth priors.InstanceLoc supports near-real-time localization in 3DIML.Source: 4) Semantic RGB SLAM · p. 14

5) Uncertainty Estimation:

  • NeRF-SLAM pairs DROID-SLAM tracking with an Instant-NGP radiance-field map and uses estimated depth uncertainty to weight the loss.Source: 5) Uncertainty Estimation · p. 14

C. LiDAR-Based SLAM Strategies

  • LiDAR-based SLAM addresses outdoor settings where RGB-D cameras are impractical, using sparse, accurate long-range depth while confronting sparse coverage and missing RGB information.The survey categorizes these methods into NeRF- and 3DGS-style approaches.Source: C. LiDAR-Based SLAM Strategies · p. 14

NeRF-LOAM [15]

  • NeRF-LOAM and later neural LiDAR systems use incremental implicit maps, with approaches adding parallel tracking and mapping, elastic point-based representations, or compact multi-scale maps.TNDF-Fusion is reported to reduce memory consumption while maintaining mapping quality from room-sized scenes to city landscapes.Source: NeRF-LOAM [15] · p. 14

2) 3DGS-style LiDAR-based SLAM:

  • 3DGS-style LiDAR methods fuse sensor streams by converting LiDAR data into Gaussian distributions or combining point-cloud registration with image refinement.LIV-GaussMap uses photometric cues to optimize Gaussian appearance and geometry; MM-Gaussian refines Gaussian attributes through keyframe processing.Source: 2) 3DGS-style LiDAR-based SLAM · p. 14

IV. EXPERIMENTS AND ANALYSIS

  • The experiment analysis compiles cross-paper quantitative results under shared evaluation protocols, prioritizing consistent, verifiable benchmarks and using one NVIDIA 3090 GPU for runtime and memory comparisons.The surveyed results span tracking, reconstruction, rendering, semantic segmentation, runtime, and memory.Source: §IV · pp. 14–15

A. Visual SLAM Evaluation

  • Visual SLAM evaluation covers RGB-D and RGB tracking, reconstruction, rendering, and runtime or memory, with semantic segmentation assessed on Replica because it provides ground-truth labels.The named datasets are TUM-RGB-D, Replica, and ScanNet.Source: A. Visual SLAM Evaluation · p. 15

1) Tracking:

Tracking results vary by sensor setting: radiance-field methods often trail established RGB-D baselines, while systems with external trackers and additional optimization are stronger. Mapping and rendering results show improvements among newer methods, with outcomes depending on the task and representation.

1) Tracking:

  • Across TUM RGB-D, ScanNet, and Replica, RGB-D radiance-field systems generally trail BAD-SLAM and ORB-SLAM2, while external tracking with Global BA or loop closure distinguishes stronger methods.GO-SLAM, Loopy-SLAM, and MoD-SLAM stand out on Replica; ScanNet results identify MoD-SLAM and GO-SLAM among the leading RGB-D methods.Source: 1) Tracking · pp. 15–16, Table IV · p. 16
  • For RGB tracking, ORB-SLAM2 is more accurate than DROID-SLAM as a baseline, while Orbeez-SLAM, MoD-SLAM, and MonoGS++ reach ATE RMSE comparable to the best RGB-D methods using external trackers.The RGB-D methods without additional tracking or optimization strategies remain more vulnerable to trajectory drift and noise.Source: 1) Tracking · p. 15
  • On ScanNet, MoD-SLAM achieves the best average ATE RMSE at 6.23, while SLAIM records 6.32 without an additional tracker; the comparison also includes qualitative trajectory results.The ScanNet tracking table reports ATE RMSE in centimeters; the figure visualizes ground-truth and estimated trajectories and ATE.Source: 1) Tracking · p. 15, Table III · p. 15, Figure 10 · p. 16

2) Mapping:

  • On Replica mapping, CG-SLAM achieves the best overall score of 1.01, while Point-SLAM reaches 1.41 in Accuracy; Loopy-SLAM leads the L1-Depth metric.The cited table evaluates L1-Depth, accuracy, completion, and completion ratio.Source: 2) Mapping · pp. 16–17, Table V · p. 16
  • For RGB mapping on Replica, MoD-SLAM is most accurate, MGS-SLAM has the highest completion, and Q-SLAM leads L1-Depth at 2.76; RGB-only methods generally fare less well than depth-based methods.The distinction among RGB methods is less pronounced than in RGB-D results, with iMAP the stated exception to the RGB-versus-depth pattern.Source: 2) Mapping · p. 17
  • On Replica rendering, Gaussian-splatting and neural-point RGB-D methods outperform earlier neural SLAM methods across PSNR, SSIM, and LPIPS, with over 10 dB improvement in PSNR; GS-ICP SLAM averages 38.83 PSNR.The qualitative rendering comparison reports improved fine-detail rendering and identifies GS-SLAM as having superior rendering quality due to its 3DGS representation.Source: 2) Mapping · p. 17, Table VI · p. 17

3) Image Rendering:

Across the reported benchmarks, PIN-SLAM outperforms LiDAR SLAM systems on KITTI, while SemGauss-SLAM has the highest Replica input-view semantic mIoU; the Replica rendering evaluation may be biased by testing on its training views.

3) Image Rendering:

  • Replica rendering scores on the same views used for training may be biased by high model capacity and potential overfitting.The analysis supports exploring alternative novel-view rendering evaluations because current SLAM benchmarks have limitations.Source: 3) Image Rendering · p. 17

4) Semantic Segmentation Results:

1) Tracking:

  • PIN-SLAM averages 1.0 m ATE RMSE across all eleven KITTI sequences and 1.1 m on sequences with loop closure, outperforming state-of-the-art LiDAR SLAM systems.PIN-LO results also underscore PIN-SLAM's improvement in global trajectory consistency.Source: Table VIII · p. 18, 1) Tracking · p. 18
  • PIN-LO achieves 0.5% translation error on KITTI odometry, competing with CT-ICP and outperforming Nerf-LOAM.PIN-LO disables loop-closure correction and pose-graph optimization; the comparison attributes its performance against Nerf-LOAM to improved SDF training and robust point-to-SDF registration.Source: Table VIII · p. 18, 1) Tracking · p. 18
  • On Newer College, PIN-SLAM's average ATE RMSE is 0.19 cm across sequences, five times lower than PIN-LO.The passage presents this as evidence of stronger global trajectory tracking.Source: Table IX · p. 18, 1) Tracking · p. 18
  • On Newer College's Quad and Math Institute sequences, PIN-SLAM is reported as the best LiDAR-based SLAM system for reconstruction quality.Accuracy and completeness are the evaluation metrics.Source: Table X · p. 18, 1) Tracking · p. 18

2) Mapping:

The reviewed mapping results compare reconstruction completeness, while the broader analysis finds that radiance-field SLAM methods still face trade-offs in runtime, memory, representation, and robustness. The discussion also identifies open challenges in global optimization, evaluation, dynamic scenes, and long-term mapping.

2) Mapping:

  • On Quad, PIN-SLAM’s completeness margin over Nerf-LOAM is about 7 cm.Source: 2) Mapping · p. 18

C. Performance Analysis

  • PIN-SLAM uses 7 GB of GPU memory at nearly 7 FPS, versus Nerf-LOAM’s nearly 12 GB and 4 seconds per frame; most reviewed systems remain short of real-time requirements.The benchmark measures GPU memory and average FPS for Replica RGB/RGB-D systems and KITTI 00 LiDAR systems.Source: C. Performance Analysis · pp. 18–19, Table XI · p. 19
  • NeRF-SLAM is 6× faster than DIM-SLAM, while hierarchical feature grids use 4 to 5× less GPU memory than SplaTAM.Source: C. Performance Analysis · p. 19

V. DISCUSSION

  • Point-based representations concentrate density around surfaces and support local updates, while 3DGS offers fast rendering but faces growth, geometry, scalability, and memory limits.Point-based methods show superior 3D reconstruction on Replica, but volumetric ray sampling restricts their efficiency.Source: §V · pp. 19–20
  • Catastrophic forgetting and submap fusion remain challenges, while current 3DGS systems struggle to process tracking and mapping together in real time.Sparse-frame sampling can produce slower, less uniform updates than sparse-ray sampling.Source: §V · pp. 19–20
  • Loop closure can require grid reallocation and map retraining, while decoupled tracking and mapping tends to improve tracking at increased computational overhead.As frames accumulate, loop-correction challenges can contribute to camera drift, inconsistent reconstruction, or rapid reconstruction collapse.Source: §V · p. 20
  • Inconsistent benchmarks hinder fair comparisons; the survey calls for standardized protocols and reliable ground truth for tracking, mapping, and novel-view rendering.Source: §V · pp. 20–21
  • Static-scene assumptions, sensor noise, non-Lambertian surfaces, RGB-only depth ambiguity, and view-dependent appearance challenge reconstruction, tracking, or rendering quality.Source: §V · p. 21

VI. CONCLUSION

The survey reviews SLAM methods influenced by radiance-field representations, from early work to recent advances. It classifies and compares methods across tracking, mapping, and rendering, and identifies limitations and open challenges.

VI. CONCLUSION

  • The survey organizes radiance-field-inspired SLAM advances and comparative results across tracking, mapping, and rendering, while identifying limitations and future research challenges.Source: §VI · p. 21
Loading 2402.13255v3…