Star 历史趋势
数据来源: GitHub API · 生成自 Stargazers.cn
README.md

Awesome Transformer-based SLAM

This repository contains a curated list of resources addressing SLAM-related tasks employing Transformer, including optical flow, view/feature correspondences, stereo matching, depth estimation, 3D reconstruction, pose estimation, etc.

If you find some ignored papers, feel free to create pull requests, or open issues.

Contributions in any form to make this list more comprehensive are welcome.

If you find this repository is useful, a simple star (Github stars) should be the best affirmation. 😊

Feel free to share this list with others!

Overview

Transformer-based SLAM

Full SLAM, including pose and depth

YearVenuePaper TitleRepositoryNote
2026arXivF²SLAM: Turning Feed-Forward Geometry into Persistent Factors for SLAM---现有做法把前馈几何当成外部状态:在局部窗口或子图里预测位姿和深度,再与在线估计做一次对齐或融合。对齐之后,这些多视图证据不在稠密光束法平差的测量里,也不能和后续局部跟踪一起重新加权。F²SLAM 只保留一套状态:关键帧位姿、稠密逆深度,以及可选的内参。因子图的每条边存的是源像素在目标关键帧上的稠密目标坐标和置信度,BA 让当前几何的重投影去贴这些目标。高频流用 DROID 式相关体积跟踪选择关键帧、维护边,并把目标写成当前重投影加一个修正。低频流在图像运动攒够的窗口里,用 DINO 特征和多视图 Transformer(Flow3r)预测整段位移和置信度,只尝试改写窗口内已有边的测量,不加新边。改写前用当前位姿和逆深度做一致性检查:支持像素够多、内点比例和残差组成的分数够高才替换,并锁住该边,避免马上被高频跟踪盖掉;否则原测量不动。前馈位姿和深度在对齐到当前尺度与参考系后只作初值,之后的细化都在这同一次 BA 里。
2026ECCVBLASt3R: Bundle Adjustment of Any Image Set with Multi-View Matching and Monocular PriorsGithub starswebsite
用学习的方法做匹配和初始化,用优化的方法做精化和全局一致。这不是简单的“先学习后优化”流水线,而是匹配、可调单目深度和光束法平差三者紧密耦合的统一框架,在线 SLAM 和无序离线重建共用同一套优化和同一组超参,默认不给内参。匹配网络用因果记忆产出跨视图像素轨迹,既用来初始化稀疏三维点,也构成重投影约束。单目网络为每张图预测针孔射线和多通道对数深度;深度不是冻结先验,而是少量系数对这组深度基底的线性组合,系数本身与焦距、位姿、三维点一起优化。残差同时包含轨迹的像素重投影和对数深度,对应关系和单目几何互相拉扯。深度项的权重先取得很小,用来稳住相机;相机固定后再加大,把各视图的稠密点图缝成全局一致的几何。序列变长后只对检索到的约 40 帧做匹配;在线时新帧先做局部优化,每隔一段再做全局优化。
2026ECCVUniSim-SLAM: Feed-Forward SLAM with Unified Sim(3) OptimizationGithub starswebsite
前馈几何随这次送进模型的视图而变:同一帧在双视图和多视图里,尺度和局部坐标系都不一样,沿长序列串起来会漂。只做双视图跟踪延迟低,但约束少;只做多视图子图更稳,却要攒够一个窗口,窗口不重叠时校正也传不下去。UniSim-SLAM 两路一起用。前端对相邻关键帧做双视图推理,在线串成全局轨迹;后端按关键帧窗口做多视图子图(默认都用 VGGT)。全局关键帧位姿和子图位姿放进同一张 Sim(3) 因子图:时序边把整段轨迹连起来,校正可以跨过不重叠的子图;视图–子图边把局部位姿接到全局轨迹,并用两套深度的中值比值锚定相对尺度;重叠子图之间再直接约束位姿和尺度,避免只靠中间帧把对齐凑出来。回环检索通过后,把查询帧和匹配帧周围的帧合成一个回环子图,写进同一张图。
2026arXivGlob3R: Global Structure-from-Motion with 3D Foundation ModelsGithub starswebsite
前馈几何模型(VGGT / Pi3X)位姿与尺度只是粗的,切块再 Sim(3) 对齐约束少、会漂;经典全局 SfM(GLOMAP)准但位姿图一脏就脆。Glob3R 用基础模型出稳健先验和跨视图特征,把它们变成可优化的多视图轨迹,再走全局 SfM(基础模型与全局几何优化相结合):冻结 Pi3X,加上稠密 warp 匹配头,把窗内关键帧到邻帧的 warp 抽成稀疏多视图轨迹;半重叠滑窗把轨迹与相对位姿传到整段(无序图用 SALAD 排成伪序列)。再按帧做旋转/平移平均与 BA,而不是只在块边界拼接。稠密深度用优化后的稀疏点做 RANSAC 尺度对齐后反投影融合。
2026arXivSLAMFormer-∞: Infinite SLAM Transformer for Unbounded Frontend and Backend ProcessingGithub starswebsite
SLAMFormer-∞ 在 SLAM-Former 上做无界前后端:不用第一帧锚定,而用 memory condition 给当前块定义局部坐标系与尺度。前端在局部窗口内用 KV cache 做在线跟踪与建图;每若干关键帧做局部后端;回环或序列结束时用同一 Transformer 做 PGGO,在位姿–几何图上迭代联合更新位姿与稠密点图(点图直接替换,位姿做阻尼插值)。骨干与 SLAM-Former 相同,训练时四种注意力掩码共用权重。室外 PGGO 初值来自与 VGGT-Long 同类的位姿图。
2026ACM MMVGGT-Align: Bridging Local Reconstruction and Global Consistency for Long-Sequence 3D ReconstructionGithub stars实现基于 VGGT-Long 的切块–Sim(3)–回环流水线。针对重叠区 IRLS 尺度无约束、误差沿链相乘(轨迹收缩或块边界错位),提出 SGIA:从各块预测点云提取相机离地高度与路面宽度等跨块不变量,用测量比作为相对尺度替换 IRLS 尺度,将 7 自由度 Sim(3) 锁成 SE(3);提取失败则退回 IRLS。另有轻量 TTA:只微调 LayerNorm,用光度/几何/平滑自监督改善后续块。两模块即插即用、无需离线重训。长回环上的旋转漂移无法仅靠尺度锚定消除。
2026RALScaRF-SLAM: Scale-Consistent Reconstruction with Feed-Forward Models and Classical Visual SLAMGithub starswebsite
解耦式稠密视觉 SLAM:经典特征 SLAM(ORB-SLAM3 / OpenVINS 等)负责低延迟跟踪,几何基础模型(Depth Anything 3)仅用于建图,避免将 GFM 几何误差传入位姿估计。将带位姿的关键帧分批送入 DA3(可输入内参与位姿),每个 batch 构成子地图;帧级用 LightGlue 稀疏匹配与 GTSAM 优化每帧尺度,再按投影深度一致性做置信度加权点云融合;子地图间以重叠帧建立对应,仅优化子地图尺度(位姿锚定于 SLAM,不对 GFM 预测做 Sim(3)/SL(4) 对齐)。回环后按更新轨迹变换子地图并做全局尺度优化。关键帧按位姿基线自适应选取(开阔场景增大平移阈值),并按中值深度剔除远距不可靠点。兼容单目/双目/视觉惯性、多相机与鱼眼(跟踪用原图,建图用针孔矫正)。建图约 3 Hz;室内每 10 m 轨迹重建误差约 2 cm,室外每 30 m 约 10 cm(以 LiDAR 点云为真值)。
2026CVPRVGGT-SLAM++---在VGGT-SLAM基础上构建完整视觉SLAM:前端用VGGT生成子地图并用Sim(3)做时序对齐;针对原方法过度依赖稀疏回环、短时漂移大、平面场景不稳的问题,提出DEM(Digital Elevation Map,数字高程图)增强的子地图表征——将VGGT稠密点云投影为平面规范的2.5D DEM并切成2×2m tile,用DINOv2提取结构感知特征、FAISS-HNSW检索构建共视图,再在共视窗口内用AnyLoc做VPR回环;后端以高频率局部空间校正/Sim(3)优化抑制漂移,实现有界内存的大规模建图。
2026arXivSceneVGGT: VGGT-based online 3D semantic SLAM for indoor scene understanding and navigationGithub starswebsite
采用YOLOv9e对2D RGB进行分割,用VGGT实现2D语义到3D到映射;采用滑动窗口方法处理连续视频流,将流划分为大小为 n 的连续不相交块,每个窗口包含当前块和前一个块的部分帧作为关键帧锚点。通过这些重叠的锚帧估计块间的位姿变换,使当前块的相对变换与先前累积的轨迹对齐,确保连续子地图与初始全局锚点对齐。将VGGT预测的深度与LiDAR到深度进行对齐获取绝对尺度。RTX 4090 上速度为7.23fps
2025arXivDynamic Visual SLAM using a General 3D PriorGithub stars采用patch-based bundle adjustment(DPV-SLAM)与前馈模型(π³mos)相结合的方式实现SLAM,通过前馈模型来过滤动态区域,然后用深度预测来进行inpainting;π³mos是基于π³的,先通过DINOv2来提取语义(移动物体);对于每帧的输入,π³预测pixel motion以及深度图,然后走dpv-slam的流程实现定位与建图
2026CVPRZipMap: Linear-Time Stateful 3D Reconstruction with Test-Time TrainingGithub starswebsite
采用 Test-Time Training (TTT) layers 作为全局信息聚合的核心机制:并非对所有 token 执行全局自注意力,而是通过 TTT 将输入视图的上下文写入一组 fast weights(动态参数),并在测试时通过参数更新过程隐式构建一个可查询的 scene state。这种方式本质上是一种 状态化建模(stateful modeling),能够在保持跨视图全局一致性的同时,将计算复杂度从二次降低到线性(即在一次前向过程中,将大规模图像集合“压缩(zip)”为一个紧凑的隐藏场景状态)。在架构上,模型结合 局部窗口 attention + TTT layers,逐步将多视图信息压缩为一个紧凑的 scene representation,并支持后续的几何与新视角查询。ATE和runtime都远超VGGT、Pi3、CUT3R和TTT3R
2025arXivVGGT-SLAM 2.0: Real time Dense Feed-forward Scene ReconstructionGithub starsTest
VGGT-SLAM的高维15自由度对齐会引入漂移,通过强制规定子地图对齐中的两个重叠帧必须具有相同的位置、旋转和标定参数,并求解一致的尺度因子,消除了VGGT子地图对齐的高维漂移与平面退化问题;改进因子图设计(所有关键帧均为节点,图中包含连接子地图内关键帧的内部边,以及连接子地图间重叠帧的中间边/其实也即是正常的因子图);VGGT注意力层的输出作为图像匹配(建立有效的回环检测),建图完成后可用于检索开放集物体,在Jetson Thor上实现在线实时建图
2025arXivDVGT: Driving Visual Geometry TransformerGithub starswebsite
视觉特征提取部分采用DINOv3作为骨干网络提取视觉特征,然后基于一系列注意力机制(分解式时空注意力/Factorized Spatial-Temporal Attention)实现时空几何关系的理解,最终并行的head输出全局3D点图+Ego Pose
2025arXivDriveVGGT: Visual Geometry Transformer for Autonomous Driving---面向自动驾驶的,多相机,VGGT重建
2025arXivAny4D: Unified Feed-Forward Metric 4D ReconstructionGithub starswebsite
2025arXivBuilding temporally coherent 3D maps with VGGT for memory-efficient Semantic SLAM------
2025arXivSING3R-SLAM: Submap-based Indoor Monocular Gaussian SLAM with 3D Reconstruction Priors---室内单目Gaussian SLAM:用CUT3R等3D先验按子地图预测局部点云/位姿,经Sim(3)对齐融合进Global Gaussian Map作为可微全局记忆;再以可微渲染做逐帧LBA与离线GBA,联合优化位姿与几何以修正先验误差;并结合基于几何特征/重投影重叠的回环与子地图位姿图优化,获得紧凑一致地图,支持稠密重建与新视角合成。
2025arXivMulti-Agent Monocular Dense SLAM With 3D Reconstruction Priors------
2025arXivAMB3R: Accurate Feed-forward Metric-scale 3D Reconstruction with BackendGithub starswebsite
AMB3R 整体采用了冻结的 VGGT 作为前端来输出 pointmap 及其对应的特征,将输出的pointmap先聚合成稀疏体素网格(Sparse voxel grid)。随后,通过空间填充曲线(Space-filling curves)将这些稀疏体素序列化为一维特征序列。利用 Transformer 高效地在紧凑的体素空间中处理这些一维序列特征,随后利用 KNN 插值将处理后的特征重新映射回 2D 视角。
2025arXivLiDAR-VGGT: Cross-Modal Coarse-to-Fine Fusion for Globally Consistent and Metric-Scale Dense MappingGithub stars---
2025arXivSAIL-Recon: Large SfM by Augmenting Scene Regression with LocalizationGithub starswebsite
2025arXivEC3R-SLAM: Efficient and Consistent Monocular Dense SLAM with Feed-Forward 3D ReconstructionGithub starsWebsite
免标定单目稠密SLAM:轻量特征跟踪(XFeat+局部稀疏地图PnP)选关键帧,仅把少量关键帧(约5帧)送入VGGT/Fast3R前馈建局部子地图并估计内参;用Sim(3)点云配准接入位姿图,并以稠密重建结果回写校正前端稀疏点;同时结合局部回环(近邻投影)与全局回环(重建embedding相似度),在Sim(3)位姿图上优化,实现多视图一致、低显存(约<10GB)与实时(约30+FPS)。
2025arXivGRS-SLAM3R: Real-Time Dense SLAM with Gated Recurrent State------
2026ICRAPROFusion: Robust and Accurate Dense Reconstruction via Camera Pose Regression and OptimizationGithub stars相机位姿回归(SLAM3R+Reloc3r)提供鲁棒性,传统优化算法确保准确度。二者相结合,形成了一个简单高效的相机跟踪和场景重建系统。
2025arXivSLAM-Former: Putting SLAM into One TransformerGithub starswebsite
20263DVViSTA-SLAM: Visual SLAM with Symmetric Two-view AssociationGithub starswebsite
单目稠密SLAM:针对非对称两视图把点图都回归到第一视角坐标系、两帧几何难拆成独立节点做位姿图的问题,前端提出轻量Symmetric Two-view Association/对称双视图关联(STA)——共享权重对称encoder-decoder,对任意两帧(邻帧/回环对)各自回归局部坐标系pointmap与相对位姿(模型约仅VGGT的35%);训练用pointmap/相对位姿(含cycle consistency)与几何一致性损失。后端构建灵活Sim(3)位姿图:每视角多节点,pose edge连同一次前向的两视图、scale edge处理跨前向尺度不一致,并结合BoW回环;LM优化后输出轨迹与稠密点云。
2025arXivVGGT-Long: Chunk it, Loop it, Align it – Pushing VGGT’s Limits on Kilometer-scale Long RGB SequencesGithub stars针对室外大场景的VGGT,提出一个chunk-and-align pipeline (也就是每部分为一个VGGT,VGGT结果之间通过sim3对齐),回环检测(通过DINOv2训练的Visual Place Recognition预训练),也是基于sim3进行对齐。最后全局通过基于sim3的LM-优化;其观点是:不必再包一层很重的 BA/SLAM backend;用「切块 + 对齐 + 轻量回环」就够。
2025arXivDINO-SLAM: DINO-informed RGB-D SLAM for Neural Implicit and Explicit Representations------
2025NeurIPSVGGT-SLAM: Dense RGB SLAM Optimized on the SL(4) ManifoldGithub stars---
2025arXiv3R-GS: Best Practice in Optimizing Camera Poses Along with 3DGS---website
MASt3R-SfM+3DGS
2025CVPRAnyCam: Learning to Recover Camera Poses and Intrinsics from Casual VideosGithub starswebsite
2025CVPRSLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB VideosGithub starstest
2025CVPRMASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction PriorsGithub starsWebsite
Test
2022ECCVJperceiver: Joint perception network for depth, pose and layout estimation in driving scenesGithub stars---

Transformer-based Pose Tracking

or pose/state estimation

YearVenuePaper TitleRepositoryNote
2026CVPRKV-Tracker: Real-Time Pose Tracking with TransformersGithub starswebsite
无需训练的实时位姿跟踪:用 π³ 在关键帧上做完整双向全局注意力建图,缓存各层全局注意力的 Key-Value 作为场景表征;新帧只做自身帧内注意力,再对冻结 KV 做交叉注意力完成重定位,避免每来一帧就对全部历史重算。跟踪时不改写缓存,减轻漂移与灾难性遗忘;仅在插入关键帧时刷新 KV。场景级按间隔选关键帧,物体级按方位角/仰角变化选帧并用低置信度回退。可零样本在线物体跟踪与重建(无需深度或 CAD),跟踪约 27 FPS,相对全量重算约 15× 加速。
2026CVPRDark3R: Learning Structure from Motion in the DarkGithub starswebsite
用于暗光环境下SfM的端到端框架,将近期3D基础模型(如MASt3R)学习到的强先验知识适应到弱光环境.受师生知识蒸馏的启发,我们开发了一种训练策略,该策略将在良好曝光的原始图像对上由MASt3R预测的密集特征图,与在弱光原始图像对上由学生模型产生的特征图对齐
2026ISPRS JPRSMASt3R-Fusion: Integrating Feed-Forward Visual Model with IMU, GNSS for High-Functionality SLAMGithub stars基于MASt3R的多传感器紧耦合SLAM:用前馈pointmap回归构建Sim(3)视觉对齐约束(Hessian形式),经同构变换接入度量尺度SE(3)因子图,与IMU预积分、GNSS紧融合;分层因子图支持实时滑窗优化与含激进回环/GNSS的全局优化,实现实时位姿跟踪、度量尺度稠密感知与全局一致建图(约8GB显存可跑任意长序列)。
2025CVPRReloc3r: Large-scale training of relative camera pose regression for generalizable, fast, and accurate visual localizationGithub stars---
2025arXivDense-depth map guided deep Lidar-Visual Odometry with Sparse Point Clouds and Images------
2025RALDINO-VO: A Feature-Based Visual Odometry Leveraging a Visual Foundation Model------
2025TROBotVIO: A Lightweight Transformer-Based Visual-Inertial Odometry for RoboticsGithub stars---
2025IROSSplatPose: Geometry-Aware 6-DoF Pose Estimation from Single RGB Image via 3D Gaussian Splatting------
2025CVPRZeroVO: Visual Odometry with Minimal Assumptions---website
2025arXivXIRVIO: Critic-guided Iterative Refinement for Visual-Inertial Odometry with Explainable Adaptive Weighting------
2025IEEE AcessTransformer-based model for monocular visual odometry: a video understanding approachGithub stars---
2025arXivLight3R-SfM: Towards Feed-forward Structure-from-Motion------
2024arXivMASt3R-SfM: a Fully-Integrated Solution for Unconstrained Structure-from-MotionGithub starsMASt3R sfm version
2024CVPRVGGSfM: Visual Geometry Grounded Deep Structure From MotionGithub starswebsite
2024Msc ThesisEnd-to-End Learned Visual Odometry Based on Vision Transformer------
2024Information FusionSelf-supervised multi-frame depth estimation with visual-inertial pose transformer and monocular guidance------
2024arXivCausal Transformer for Fusion and Pose Estimation in Deep Visual Inertial OdometryGithub stars---
2024IJRADDETR-SLAM: A Transformer-Based Approach to Pose Optimization in Dynamic Environments------
2023International Conference on Neural Computing for Advanced ApplicationsA lightweight sensor fusion for neural visual inertial odometry------
2023ITM Web of ConferencesViT VO-A Visual Odometry technique Using CNN-Transformer Hybrid Architecture------
2023arXivTransFusionOdom: interpretable transformer-based LiDAR-inertial fusion odometry estimationGithub stars---
2023CVPRModality-invariant Visual Odometry for Embodied VisionGithub starsWebsite
2023MAVViTVO: Vision Transformer based Visual Odometry with Attention Supervision------
2023International Conference on Haptics and Virtual RealityVIOFormer: Advancing Monocular Visual-Inertial Odometry Through Transformer-Based Fusion------
2022IEEE Intelligent Vehicles SymposiumAttention guided unsupervised learning of monocular visual-inertial odometry------
2022IEEE-SJEma-vio: Deep visual–inertial odometry with external memory attention------
2022IROSAFT-VO: Asynchronous fusion transformers for multi-view visual odometry estimation------
2022arXivDense prediction transformer for scale estimation in monocular visual odometry------
2021Neural Computing and ApplicationsTransformer guided geometry model for flow-based unsupervised visual odometry------

Transformer-based Optical Flow

YearVenuePaper TitleRepositoryNote
2024ECCVCotracker: It is better to track togetherGithub stars---
2023arXivWin-win: Training high-resolution vision transformers from two windows------
2023arXivFlowformer: A transformer architecture and its masked cost volume autoencoding for optical flow------
2023CVPRFlowFormer++: Masked Cost Volume Autoencoding for Pretraining Optical Flow EstimationGithub stars---
2023CVPRTransflow: Transformer as flow learner------
2023ICCVCroco v2: Improved cross-view completion pre-training for stereo matching and optical flowGithub starsCross view Match
2022NIPSCroco: Self-supervised pre-training for 3d vision tasks by cross-view completionGithub starsCross view Match
2023PAMIUnifying flow, stereo and depth estimationGithub stars---
2022CVPRGmflow: Learning optical flow via global matchingGithub stars---
2022CVPRCraft: Cross-attentional flow transformer for robust optical flowGithub stars---
2022CVPRLearning optical flow with kernel patch attentionGithub stars---
2022CVPRGlobal Matching with Overlapping Attention for Optical Flow EstimationGithub stars---
2022CVPRFlowformer: A transformer architecture for optical flowGithub stars---

Transformer-based View Matching

or Data Association, Correspondences

YearVenuePaper TitleRepositoryNote
2025arXivTAPIP3D: Tracking Any Point in Persistent 3D GeometryGithub starswebsite
2025CVPRDEFOM-Stereo: Depth Foundation Model Based Stereo MatchingGithub starswebsite
depth anything v2 + RAFT-Stereo
2025arXivMatchAnything: Universal Cross-Modality Image Matching with Large-Scale Pre-TrainingGithub starswebsite
2025ICCVPOMATO: Marrying Pointmap Matching with Temporal Motions for Dynamic 3D ReconstructionGithub stars基于MonST3R,提出pointmap matching head在3D空间显式建模跨视角对应,并结合temporal motion module学习时序运动;无需外挂光流即可估计动态掩码,支持视频深度、3D点跟踪与动态重建
2025arXivCoMatcher: Multi-View Collaborative Feature Matching------
2025arXivCoMatch: Dynamic Covisibility-Aware Transformer for Bilateral Subpixel-Level Semi-Dense Image Matching------
2025CVPRTracktention: Leveraging Point Tracking to Attend Videos Faster and BetterGithub starswebsite
2025CVPRFG2:Fine-Grained Cross-View Localization by Fine-Grained Feature MatchingGithub stars---
2025arXivColabSfM: Collaborative Structure-from-Motion by Point Cloud RegistrationGithub stars---
2025arXivLoop Closure from Two Views: Revisiting PGO for Scalable Trajectory Estimation through Monocular Priors---MASt3R for Loop Closure
2025arXivSpeedy MASt3R------
2025CVPRVGGT: Visual Geometry Grounded TransformerGithub starswebsite
Test
2024ECCVGrounding Image Matching in 3D with MASt3RGithub starsWebsite
Test
2024CVPREfficient LoFTR: Semi-dense local feature matching with sparse-like speedGithub starsELoFTR
2024CVPRRoMa: Robust dense feature matchingGithub stars---
2023CVPRRotation-invariant transformer for point cloud matchingGithub starsPoint Cloud Matching
2022ECCVAspanformer: Detector-free image matching with adaptive span transformer------
2022ICARMTlcd: A transformer based loop closure detection for robotic visual slam------
2021ICCVCotr: Correspondence transformer for matching across imagesGithub stars---
2021CVPRLoFTR: Detector-free local feature matching with transformersGithub stars---
2020CVPRSuperglue: Learning feature matching with graph neural networksGithub starsborrows the self-attention

Transformer-based Mapping

or depth estimation or 3D reconstruction

YearVenuePaper TitleRepositoryNote
2026ICML4RC: 4D Reconstruction via Conditional Querying Anytime and AnywhereGithub starswebsite
4RC 从单目视频编码一次,再按需查询任意源帧到任意目标时刻的稠密三维几何与运动:每帧 4D 属性分解为视点不变的基准几何,加上随时间的三维位移。网络由一条 ViT 编码器(DINOv2 / DA3 初始化,交替帧内与全局自注意力,每帧带 camera token 与 time token)和条件解码器组成:几何头(dual-DPT + 轻量相机头)出深度、射线与相机;运动头(自注意力 + 交叉注意力,AdaLN 注入目标时刻)出位移场。另有流式变体 S-4RC,把编码器换成 STream3R 因果骨干。
2026arXivSM4RT: Learning Structured Motion Geometry for 4D ReconstructionGithub starswebsite
SM4RT 从单目 RGB 视频单次前向同时重建三维几何与结构化运动:用 Structure-of-Motion 把场景运动分解成一组 se(3) 扭转基(每条基是随时间的 6D twist 序列)和跨时间共享的逐像素归属,加权后再 Exp 到 SE(3),得到稠密世界坐标轨迹。网络由两条并行支路组成:场景几何支路沿用 DA3(DINOv2 骨干 + Camera / Depth Head)输出深度与相机;运动几何支路用运动 token 交叉注意中间几何特征,Base Head(DPT)出归属、Motion Head 出扭转基。
2026arXivIGGT4D: Streaming 4D Instance-Grounded Geometry TransformerGithub starswebsite
流式几何模型(CUT3R / StreamVGGT / Stream3R / LingBot-Map)只维持深度与位姿,不维持物体身份;实例感知模型(IGGT)能做几何+实例,但是固定图像集合 + 全局注意力 + HDBSCAN,长视频会爆显存。IGGT4D 把 DA3 改成因果流式,并加上几何锚定的实例头与常数时间在线聚类。这里的 4D 即 3D 几何(深度、射线、相机)加上跨帧持久的物体实例。按帧处理、跨视图只看历史并用 KV cache;Tri-DPT 同时解深度、射线与实例特征,再用全局码本把帧内聚类匹配成持久 ID。另构建 InsScene4D-147K 提供时间一致的实例监督。
2026ECCVReViV: Reconstructing the Viewer and the View in 4D from Monocular Egocentric VideoGithub starswebsite
从单目第一人称 RGB 同时重建观察者(全身、手、视线)与场景(仿射不变深度、相机轨迹)。在 EgoM2P / 4M 的任意到任意掩码生成上扩展:各模态用专用 VQ-VAE 离散化(身体与手为双流球面码本 tokenizer,RGB/深度用 Cosmos,视线与相机在放大数据上重训),另加 ViT 编码原视频以补量化丢失的细节。T5 式编解码器在随机多模态掩码下学联合分布;推理只给 RGB,对其余 token 做 3 步并行解码。深度无绝对尺度,用脚部高度与场景最低点拟合地面求尺度,看不见地面时可用 VIPE 度量深度作锚。数据引擎把 EgoM2P 语料扩到含手/全身的 7B unique token(Video Depth Anything 补深度伪标签;手在相机系、身体在重力对齐世界系)。不依赖预计算 SLAM 或外接手跟踪。量化会损失深度高频细节。
2026ECCVCasaMaestro: Multi-View Panoramas for House-Scale 3D ReconstructionGithub starswebsite
面向整屋的稀疏全景前馈度量重建:输入 20–50 张无外参室内 ERP 全景,一次前向同时预测各视角度量深度与相机位姿,反投影得到覆盖多房间的点云。骨干沿用 Depth Anything 3 的极简 ViT(DINOv2):前若干层只做帧内自注意力,之后通过重排 token 交替帧内/跨视图全局注意力(不加新层),DPT 出深度,位姿头吃骨干最终特征。针对稀疏大位移,不用每视角独立 MLP 回归位姿,而加一层轻量全景位姿解码器:线性投影后先做逐视角残差 MLP,再在视角维做自注意力(可用 mask 挡住三脚架/拍摄者),分别回归平移与单位四元数;全景视场固定为 360°×180°,不估 FoV。现有室内全景同场景视角几乎同朝向、相对旋转不足,故对合成场景做 ERP 旋转增强:按 yaw–pitch–roll 采样旋转,球面重映射 RGB 与深度,并右乘更新相机到世界的旋转,几何保持一致。权重从 DA3 初始化,在 Realsee3D 上训练。相对把全景切成针孔再跑 VGGT/π³/DA3,或依赖大重叠视频流的 PanoPose/SPR,稀疏整屋的位姿与房间结构更稳。
2026arXivMoGe-3: Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric RefinementGithub starswebsite
单目度量点图的细结构 refinement:现有方法在 2D 平面解码 3D 几何,邻域由图像邻近而非 3D 邻近决定,深度不连续处特征混叠,细杆/栅栏等会被拧弯、抹平。以 MoGe-2 为 base 得到粗点图与 DINOv2 特征;Self-Guided Sparse 3D Refiner 把点图抬到 log-depth 体素壳,用稀疏 3D 卷积按三维邻域聚合(仅图像相邻且 log-depth 差小于阈值才连通),预测 log-depth 残差并迭代更新壳结构,瓶颈注入 base 的 2D 特征。默认 ViT-L、推理 3 步;相对 MoGe-2 等,局部细结构与全局几何均更好。
2026arXivGUSH3R: Everyone Everywhere All at Once as GaussiansGithub starswebsite
在冻结的 Human3R 上做前馈人–场景重建:单目视频一次前向同时输出动态人与静态场景的统一 3DGS。Human3R 提供度量点云、相机位姿与 SMPL-X;Scene Gaussian Decoder 用点云作高斯中心,DPT+CNN 预测其余高斯参数,用人检测/置信度滤掉人体区域,再按度量体素累积融合。Human Gaussian Decoder 以 SMPL-X 顶点为锚,Human Gaussian Transformer 用 human/vertex/memory token 查询图像 token,在 canonical A-pose 预测逐顶点高斯,经 LBS 变到当前姿态;每人一条 appearance memory 以保持跨帧外观。相对优化式人–场景方法和 AnySplat+LHM 事后拼接,新视角合成更一致,推理更快。
2026arXivRegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer---无需训练的 VGGT 加速:谱分析、探针与因果干预表明全局注意力分成三段——浅层(L1–L10)尚无跨视角结构,中层(L11–L18)承担跨视角对齐,深层(L19–L24)对稠密几何冗余,但跨帧注意力对位姿仍必要。据此在每层全局注意力上沿两条轴做 U 形压缩:Saliency-Guided Banded Merging 用冻结 DINOv2 [CLS] 显著度保护几何/边缘 salient token(浅/深层合并比 0.99,中层 0.50);Selectively Protected K/V Downsampling 对非锚点帧做相位错开的空间网格降采样,参考帧、camera/register token 与高显著 patch 保持满密度,以维持跨帧覆盖与位姿通路。相对 VGGT*,ScanNet 1000 帧约 6.7× 加速,重建质量基本持平。
2026arXivLiAuto-GeoX: Efficient Grounded Driving TransformerGithub starswebsite
面向车载部署的环视稠密三维重建:先以 VGGT/OmniVGGT 风格的高容量教师(24 层编解码、约 1.1B)在大规模环视数据上学习驾驶几何,训练时将标定多相机内外参作为显式几何条件,并以稀疏 LiDAR 投影深度对远距、弱纹理区域做几何锚定;再通过 geometry-preserving distillation 将能力迁移至 155M 学生(DINOv2-Small + 12 层 decoder)。蒸馏包含两项:Mask-Guided Depth-Aware Distillation 利用教师高激活 token 与深度图,经 adapter 调制学生特征,强化远距与边界等几何敏感区域(仅训练期使用,推理无额外开销);Relative-Pose Relational Distillation 对齐各视角 camera-conditioned token 的成对关系矩阵,以保持环视空间一致性,而非直接回归绝对位姿。学生以教师稠密深度与点云为监督,不以稀疏雷达为监督。推理仅部署学生,单次前向输出各视角深度与稠密点云,适配 2–8 路相机配置。KITTI 视频深度约 220 FPS。
2026arXivBA-T: An Iterative Transformer for Two-View Bundle AdjustmentGithub starsfeed-forward 3D多数是“一次性预测”,本文引入BA的思想(并非显式求解):既保留深度模型的隐式表达和前馈效率,又吸收 BA 的迭代式信息传播结构。
两视图前馈重建:把经典 BA 的「残差→更新相机→更新几何」信息流做成可复用的隐空间迭代 Transformer 层;用 camera-conditioned transform (先把不同视角的局部几何拉到可比较的空间) + soft matching 算 latent residual,再结构化更新位姿/局部几何 token,替代深层 cross-attn 堆栈。decoder 约 38M(约 DUSt3R 的 16%),迭代 3–4 次 refinement,位姿与跨视图一致性优于同体量堆栈 decoder。
2026CVPRD4RT: Efficiently Reconstructing Dynamic Scenes One D4RT at a TimeGithub stars
Github stars
Github stars
Github stars
website
面向动态视频的统一前馈 4D 重建与跟踪:Encoder 把整段视频编码成全局场景表征,轻量 Decoder 按需查询任意像素在指定时刻、指定相机坐标系下的 3D 位置;通过更换查询方式,同一模型可输出点轨迹、稠密点云、视频深度,以及相机内外参(相对位姿由多点 3D 对应经刚体对齐得到)。避开稠密逐帧解码和多任务专用头;查询彼此独立并附带局部 RGB 外观块。再用占用网格做全像素跟踪,只从尚未覆盖的像素开新轨迹,高效重建含动态物体的完整场景。相对 VGGT、π³、MegaSaM、SpatialTrackerV2 等,在 4D 跟踪、深度/点云与位姿上精度与速度均更强。
2026arXivSpark3R: Asymmetric Token Reduction Makes Fast Feed-Forward 3D Reconstruction---无需训练(即插即用),加速前馈3D重建模型,兼容VGGT、π3 、DepthAnythingV3(1000帧上高达28倍加速).将query tokens 和 key-value tokens的压缩解耦了,并分别配不同的压缩算子。ayer-adaptive key-value reduction schedule :根据每个全局注意力层对压缩的实测敏感性,调整其键值缩减因子
2026arXivTurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention---website
可训练加速 VGGT 类模型:Adaptive Alternating Attention = 按帧/层自适应稀疏选择 + 学习代表性 token 的稀疏全局 cross-attn;兼容 VGGT/π³/MapAnything,1000 帧约 7×–18× 加速
用 Adaptive Alternating Attention 替换稠密全局注意力,在保持重建质量的同时显著降时延与显存
2026CVPRFrom None to All: Self-Supervised 3D Reconstruction via Novel View SynthesisGithub starswebsite
通过自监督的方式来学习显式的3D几何和相机内外参。通过重构3DGS,然后渲染,利用2D图像的自监督
2026CVPRVGGT-ΩGithub starswebsite
将模型参数量扩展至10B,引入寄存器注意力(Register Attention):ViT往往自发利用小部分的Token携带全局信息,因此作者在 25% 的全局注意力层中,不同帧之间的信息交换被严格限制在寄存器之间,形成了有效的信息瓶颈,强制寄存器提取高层几何特征;对于解码头进行显存优化;从4000万数据中选出80万高质量序列,结合合成数据有400万个带3D label的训练序列(原本VGGT的15倍以上)。
2026ICLRTtt3r: 3d reconstruction as test-time trainingGithub starswebsite
基于CUT3R的改进,CUT3R属于RNN架构,参考现有的RNN架构来改进CUT3R;TTT正是为了解决RNN长度泛化能力(隐藏状态在训练见过的长度/模式上“记死”了;信息在时间上被不断覆盖或衰减,长序列里早期重要信息丢失;序列变长后,隐藏状态会进入训练阶段几乎没出现过的区域,行为不可控、误差容易放大)
2026CVPRSwiftVGGT: A Scalable Visual Geometry Grounded Transformer for Large-Scale ScenesGithub starswebsite
在保持高质量密集3D重建的同时显著降低了推理时间;主要改进点有二:(1)将基于迭代IRLS的Sim(3)对齐替换为基于单步Sim(3)的奇异值分解流程;(2)直接使用VGGT的DINO块标记执行闭环检测,消除了外部VPR模型引入的冗余计算(实现回环检测)
VGGT-Long 的加速版——可靠点采样换掉 IRLS,自带 DINO 特征换掉外部 VPR
2026arXivGeometric Context Transformer for Streaming 3D ReconstructionGithub starswebsite
LingBot-Map,纯粹的自回归模型,单个RGB实现流式 3D 重建;DINO 骨干网络提取图像特征,经过交替排列的 Frame Attention 层和 GCA 层进行精炼。在 GCA 中,当前视图从 Anchor Context、局部 Pose-Reference Window 和 Trajectory Memory 中聚合信息。任务特定的预测头分别输出相机位姿和深度图。
2026CVPRPanoVGGT: Feed-Forward 3D Reconstruction from Panoramic Imagery---基于全景图的前馈三维重建;引入了针对全景图像的特殊三轴SO(3)旋转处理机制,关键应该是数据集
2026ICLRPAGE-4D: Disentangled Pose and Geometry Estimation for VGGT-4D PerceptionGithub starswebsite
通过对VGGT关键层的特征可视化分析,研究团队观察到动态区域表现出比静态区域更弱的激活,这表明VGGT倾向于忽略动态内容。提出Dynamics-aware Aggregator:通过三个模块整合空间和时间线索——Frame Attention用于帧间patch关系,Global Attention用于帧内patch关系,Dynamics-Aware Global Attention用于解耦动态和静态内容;Dynamics-aware Aggregator预测动态感知掩码,实现静态与动态信息的解耦:在姿态估计中抑制运动线索,在几何重建中增强动态表示。
2026arXivMeMix: Writing Less, Remembering More for Streaming 3D ReconstructionGithub starswebsite
“流式 3D 重建”是指模型像人眼一样,不断接收连续的视频帧流,并实时构建周围的 3D 场景。为了处理无限长的视频流,现有的主流模型通常采用“循环状态(Recurrent State)”机制——即把过去的记忆压缩到一个固定大小的状态中。本文不再把状态看作一个整体,而是将其划分为多个独立的记忆补丁(patches),然后选择性的更新,这样可以避免长序列的遗忘问题。
2025arXiv4D-VGGT: A General Foundation Model with SpatioTemporal Awareness for Dynamic Scene Geometry Estimation---Multi-level representation:对于空间(Spatial),采用“跨视图的全局融合(cross-view global fusion)”,以确保同一时刻下多视角几何的全局一致性。对于时间(Temporal),采用“跨时间的局部融合(cross-time local fusion)”,以精准捕捉物体在前后帧之间的局部运动细节。在获取了时空特征的基础上,通过多个prediction head来预测深度、场景流等
2026arXivR3: 3D Reconstruction via Relative RegressionGithub starswebsite
针对 VGGT/π³/DA3 等依赖单一全局坐标系、长序列平移无界难扩展的问题,将前馈重建改为成对相对位姿回归:在 DA3 骨干上用轻量 MLP 预测有序帧对的相对旋转/平移及分项置信度;置信度统一用于训练加权、流式轨迹聚合与关键帧库管理(token 新颖性准入、效用淘汰),并可拒不可靠帧。同一因果 checkpoint 既支持有界内存在线流式,也可测试时去因果 mask 做全上下文 + 轻量位姿图 refinement。约 372M 参数,长序列位姿/重建相对 CUT3R、StreamVGGT、TTT3R 等更稳。
2026arXivFrameVGGT: Coherence-Preserving Memory for Bounded Streaming GeometryGithub stars针对 StreamVGGT 等流式几何模型 KV-cache 随序列无限增长的问题,提出推理时有界记忆框架(无需重训):主张几何匹配依赖帧级连贯、跨视角兼容的观测,token 级淘汰易导致上下文稀释与碎片化。将历史按帧保留为完整 KV 段;中期库用 key 均值原型做互补性评分,并以最远优先贪心在固定容量下覆盖更广的检索方向;另设稀疏锚点层(首帧永久保留,按时间间隔/置信清晰度/位姿新颖性晋升)提供长程参考。在长序列重建、视频深度与位姿上,相对 InfiniteVGGT、OVGGT 等 token 级有界方法取得更好的精度–显存权衡,且更稳定。
2026arXivOVGGT: O (1) Constant-Cost Streaming Visual Geometry TransformerGithub starswebsite
Self-Selective Caching (SSC,自选择性缓存) and Dynamic Anchor Protection (DAP,动态锚点保护)。固定的显存下,处理任意长度序列
2026arXivVGG-T^3: Offline Feed-Forward 3D Reconstruction at Scale---website
VGG-T³(Visual Geometry Grounded Test-Time Training):面向离线大规模重建,将VGGT全局Softmax Attention中可变长度的KV场景表征,通过测试时训练(TTT)压缩为固定尺寸MLP权重,使复杂度从O(n²)降为线性O(n);并引入ShortConv2D增强KV映射表达力。约1k图54秒、2k图约48.5秒(相对VGGT最高约33×加速),且可用冻结MLP查询新视角做视觉定位。
2026arXivFlow4R: Unifying 4D Reconstruction and Tracking with Scene Flow---website
Flow4R会对每一幅图像预测点位置(相当于pointmap)、场景流、姿态权重和置信度,其中,场景流(scene flow)实现了捕捉点相对于摄像机的运动情况,概括来说就是一个模型同时看懂场景、跟踪物体、估相机运动、完成重建”,而它选择的统一语言不是传统位姿或深度,而是 scene flow
2026CVPRLiteVGGT: Boosting Vanilla VGGT via Geometry-aware Cached Token MergingGithub starswebsite
针对VGGT全局注意力二次复杂度瓶颈,提出geometry-aware cached token merging:用像素梯度+token方差构建几何重要度图,将token划分为保留的GA tokens、锚点dst与待合并src,在全局注意力两侧做合并/还原;并跨层缓存复用merge indices。再结合微调与FP8量化,可处理约1000张图、相对VGGT约10×加速且显存显著降低。
2026arXivFast-SAM3D: 3Dfy Anything in Images but FasterGithub stars从输入图片中重构物体三维模型
2026arXivMasked Depth Modeling for Spatial PerceptionGithub starswebsite
面向真实场景的深度补全模型:将传感器失效区域视为可学习的结构线索而非简单噪声,通过掩码深度建模(Masked Depth Modeling, MDM)机制充分利用 RGB 图像中的视觉上下文信息,对缺失深度区域进行智能推断与补全。架构:RGB+Deoth输入ViT(标准架构);合成数据LingBot Depth-S(生成带有自然缺陷的逼真深度观测值,442个室内场景中渲染了100万个合成样本)+真实数据LingBot Depth-R(200万个场景)+开源数据集补充(共1000万个训练样本)
2026arXivV-DPM: 4D Video Reconstruction with Dynamic Point MapsGithub starswebsite
将动态点图(DPM)从图像对扩展为多帧视频,进而可以实现4D重建;建立于VGGT基础上,通过time-conditioned decoder将重建过程分解为两个步骤:用VGGT生成视角不变但时间可变的点图(即所有点都在同一个相机坐标系)+用时间条件解码器生成时间不变的点图(即不仅在同一个坐标系下,也传递到同一时间戳上)
2026arXivMVGGT: Multimodal Visual Geometry Grounded Transformer for Multiview 3D Referring Expression SegmentationGithub stars从稀疏的RGB图像和文本描述中,同时完成场景重建和目标分割;双分支架构:重建Branch(三维重建)和多模态Branch(接收来自于重建Branch的几何特征,与图像的视觉特征、文本encoder进行融合,实现对描述区域的关注,输出2D mask并投影到3D点云上)
2025arXivOne4D: Unified 4D Generation and Reconstruction via Decoupled LoRA ControlGithub starswebsite
一个模型(视频生成模型Wan Video)同步输出 RGB 视频与 Pointmap(XYZ)几何视频
2026CVPRTALO: Pushing 3D Vision Foundation Models Towards Globally Consistent Online ReconstructionGithub stars全局一致性的在线增量建图;基于全局控制点传播与薄板样条变换的精细几何对齐策略,基于重叠相机位姿平均的子地图注册,并支持不同的基础模型(如VGGT,π³、MapAnything)
2025arXivInfiniteVGGT: Visual Geometry Grounded Transformer for Endless StreamsGithub stars基于Key-Diversity的滚动记忆机制/Rolling Memory,实现面向无限长序列的视觉几何重建框架;算每个 Key 与均值 Key 的负余弦相似度。得分越高,代表该 Token 与平均值差异越大,包含越独特的几何信息,应予以保留
2025arXivDePT3R: Joint Dense Point Tracking and 3D Reconstruction of Dynamic Scenes in a Single Forward PassGithub stars利用多张图像实现动态场景的密集点跟踪与3D重建,提取 deep spatio-temporal features,通过dense prediction heads来回归pixel-wise 地图;VGGT+Dense Prediction Transformer(DPT,分别作为Depth Head、Point Head和Motion Head)
2025arXivDepth Anything 3: Recovering the Visual Space from Any ViewsGithub starswebsite
2025arXivHTTM: Head-wise Temporal Token Merging for Faster VGGT------
2025arXivFastVGGT: Training-Free Acceleration of Visual Geometry Transformer------
2025arXivVGGT4D: Mining Motion Cues in Visual Geometry Transformers for 4D Scene ReconstructionGithub starswebsite
2025arXivOmniVGGT: Omni-Modality Driven Visual Geometry GroundedGithub starswebsite
2025arXivHuman3R: Everyone Everywhere All at OnceGithub starswebsite
2025arXivMapAnything: Universal Feed-Forward Metric 3D ReconstructionGithub starswebsite
2025ICCVPanoSplatt3R: Leveraging Perspective Pretraining for Generalized Unposed Wide-Baseline Panorama ReconstructionGithub starswebsite
2025arXivOv3R: Open-Vocabulary Semantic 3D Reconstruction from RGB Videos------
2025ICCVLONG3R: Long Sequence Streaming 3D ReconstructionGithub starswebsite
2025arXivDens3R: A Foundation Model for 3D Geometry Prediction------
2025arXivStreamVGGT: Streaming 4D Visual Geometry TransformerGithub starswebsite
2025arXivTest3R: Learning to Reconstruct 3D at Test TimeGithub starswebsite
2025arXivPoint3R: Streaming 3D Reconstruction with Explicit Spatial Pointer MemoryGithub starswebsite
2025arXiv4DGT: Learning a 4D Gaussian Transformer Using Real-World Monocular Videos---website
2025arXivRig3R: Rig-Aware Conditioning for Learned 3D Reconstruction------
2025CVPRSAB3R: Semantic-Augmented Backbone in 3D Reconstruction---website
2025arXivStyl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and StylesGithub starswebsite
2025CVPRMonoSplat: Generalizable 3D Gaussian Splatting from Monocular Depth Foundation ModelsGithub stars---
2025arXivRecollection from Pensieve: Novel View Synthesis via Learning from Uncalibrated VideosGithub stars---
2025arXivSTORM: Spatio-Temporal Reconstruction Model for Large-Scale Outdoor ScenesGithub starswebsite
2025CVPRWDELTA: Dense Depth from Events and LiDAR using Transformer's AttentionGithub starswebsite
2025CVPRMV-DUSt3R+: Single-Stage Scene Reconstruction from Sparse Views In 2 SecondsGithub starswebsite
2025CVPRMoGe: Unlocking Accurate Monocular Geometry Estimation for Open-Domain Images with Optimal Training SupervisionGithub starswebsite
2025arXivRegist3R: Incremental Registration with Stereo Foundation Model------
2025arXivSt4RTrack: Simultaneous 4D Reconstruction and Tracking in the World---website
2025CVPRAerialMegaDepth: Learning Aerial-Ground Reconstruction and View SynthesisGithub starswebsite
2025arXivMono3R: Exploiting Monocular Cues for Geometric 3D Reconstruction------
2025CVPRMonSter: Marry Monodepth to Stereo Unleashes PowerGithub stars---
2026NeurIPSEnhancing 3D Reconstruction for Dynamic ScenesGithub starswebsite
D2USt3R/DDUSt3R:提出 Static-Dynamic Aligned Pointmaps(SDAP):静态区仍用相机位姿对齐,动态区用光流把像素对齐到同一时刻的 3D 坐标,并配合遮挡掩码与动态掩码,分开监督:(1) 只在静态区做常规 pointmap 回归;(2)在动态且非遮挡区,用光流建立跨帧 3D 对齐.把动态对应显式写进 DUSt3R 训练,从而提升动态场景重建。
2025arXivFlowR: Flowing from Sparse to Dense 3D Reconstructions---website
2025ICCVEasi3R: Estimating Disentangled Motion from DUSt3R Without TrainingGithub starswebsite
Easi3R:无需训练的即插即用适配方法。发现DUSt3R的cross-attention隐式编码了相机/物体运动信息,通过时空聚合attention(均值/方差)解耦动态区域与相机运动,得到动态分割;再在二次推理时对动态区域做attention re-weighting,抑制运动干扰,从而提升动态视频上的4D点云重建与相机位姿估计。
test
2025arXivSparseGS-W: Sparse-View 3D Gaussian Splatting in the Wild with Generative Priors---DUSt3R+Diffusion+3DGS
2025ICLRM3: 3D-Spatial Multimodal MemoryGithub starswebsite
compression & Gaussian Memory Attention
2025CVPRMVSAnywhere: Zero-Shot Multi-View StereoGithub starswebsite
2025CVPRCoMapGS: Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis---website
2025CVPRPow3R: empowering unconstrained 3D reconstruction with camera and scene priors---website
DUSt3R+multi information input
2025CVPRText-guided Sparse Voxel Pruning for Efficient 3D Visual GroundingGithub starsTSP3D
2025CVPRUniK3D: Universal Camera Monocular 3D EstimationGithub starswebsite
2025CVPRSonata: Self-Supervised Learning of Reliable Point RepresentationsGithub starswebsite
2024CVPRPoint transformer v3: Simpler faster strongerGithub stars---
2022NIPSPoint transformer v2: Grouped vector attention and partition-based poolingGithub stars---
2021ICCVPoint transformer---unofficial implementation
2025arXivDynamic Point Maps: A Versatile Representation for Dynamic 3D Reconstruction---website
Dynamic DUSt3R, DPM
2025ICLRMonST3R: A Simple Approach for Estimating Geometry in the Presence of MotionGithub starswebsite
MonST3R(Motion DUSt3R):以geometry-first方式将DUSt3R扩展到动态场景——按时间步回归pointmap,通过对少量动态/位姿/深度数据微调decoder与head(冻结encoder),使模型在无显式运动表示时也能估计运动物体几何;推理时用光流与相机诱导流差异得到动静掩码,并结合滑动窗口全局对齐,输出视频深度、相机位姿与动态点云。
Test
2025CVPRStereo4D: Learning How Things Move in 3D from Internet Stereo VideosGithub starswebsite
2025CVPRContinuous 3D Perception Model with Persistent StateGithub starswebsite
CUT3R
2025CVPRSPARS3R: Semantic Prior Alignment and Regularization for Sparse 3D ReconstructionGithub starsMASt3R+COLMAP+3DGS
2025arXivSplatVoxel: History-Aware Novel View Streaming without Temporal Training------
2025CVPRGaussTR: Foundation Model-Aligned Gaussian Transformer for Self-Supervised 3D Spatial UnderstandingGithub stars3DGS+Transformer
2025CVPRDUNE: Distilling a Universal Encoder from Heterogeneous 2D and 3D TeachersGithub starswebsite
distillation
2025arXivMUSt3R: Multi-view Network for Stereo 3D ReconstructionGithub starsmultiple views DUSt3R
2025CVPRFast3R: Towards 3D Reconstruction of 1000+ Images in One Forward PassGithub starsWebsite
Test
2024NIPSDepth anything v2Github starswebsite
2024CVPRDepth anything: Unleashing the power of large-scale unlabeled dataGithub starsWebsite
2024CVPRDeCoTR: Enhancing Depth Completion with 2D and 3D Attentions------
2024CVPRLearning to adapt clip for few-shot monocular depth estimation------
20253DV3d reconstruction with spatial memoryGithub starswebsite
Spann3R
2024CVPRDUSt3R: Geometric 3D Vision Made EasyGithub starsWebsite
Test
2024ECCVGs-lrm: Large reconstruction model for 3d gaussian splatting---website
3DGS+Transformer
2024TIPBinsFormer: Revisiting Adaptive Bins for Monocular Depth EstimationGithub stars---
2024TIPGLPanoDepth: Global-to-Local Panoramic Depth Estimation------
2023ICCVTowards zero-shot scale-aware monocular depth estimationGithub starswebsite
2023ICCVEgformer: Equirectangular geometry-biased transformer for 360 depth estimationGithub stars---
2023Machine Intelligence ResearchDepthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation------
2023CVPRLite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation------
2023CVPRCompletionFormer: Depth Completion with Convolutions and Vision TransformersGithub starswebsite
2023ICRALightweight monocular depth estimation via token-sharing transformer------
2023AAAIROIFormer: Semantic-Aware Region of Interest Transformer for Efficient Self-Supervised Monocular Depth Estimation------
2023ICRATODE-Trans: Transparent Object Depth Estimation with TransformerGithub stars---
2023AAAIDeep digging into the generalization of self-supervised monocular depth estimationGithub stars---
2022ECCVPanoFormer: Panorama Transformer for Indoor 360 Depth EstimationGithub stars---
2022AAAIImproving 360 monocular depth estimation via non-local dense prediction transformer and joint supervised and self-supervised learning------
2022arXivMVSFormer: Multi-view stereo by learning robust image features and temperature-based depth------
2022arXivObjcavit: improving monocular depth estimation using natural language models and image-object cross-attentionGithub stars---
2022arXivDepthformer: Multiscale Vision Transformer For Monocular Depth Estimation With Local Global Information FusionGithub stars---
2022arXivSidert: A real-time pure transformer architecture for single image depth estimation------
2022ECCVHybrid transformer based feature fusion for self-supervised monocular depth estimation------
2022ECCVSpike transformer: Monocular depth estimation for spiking cameraGithub stars---
20223DVMonoViT: Self-Supervised Monocular Depth Estimation with a Vision TransformerGithub stars---
2022arXivDEST: "Depth Estimation with Simplified Transformer------
2022arXivSparseFormer: Attention-based Depth Completion Network------
2022CVPRGuideFormer: Transformers for Image Guided Depth Completion------
2022CVPRMulti-frame self-supervised depth with transformers------
2022arXivTransformers in Self-Supervised Monocular Depth Estimation with Unknown Camera Intrinsics------
2021ICCVRevisiting stereo depth estimation from a sequence-to-sequence perspective with transformers---STTR
stereo matching
2021BMVCTransformer-based Monocular Depth Estimation with Attention SupervisionGithub stars---
2021ICCVTransformer-Based Attention Networks for Continuous Pixel-Wise PredictionGithub stars---
2021ICCVVision transformers for dense predictionGithub starsDPT

Other Resources

YearVenuePaper TitleRepositoryNote
2025arXivDINOV3Github starswebsite
2024Transactions on Machine Learning Research JournalDinov2: Learning robust visual features without supervisionGithub starsDINO2
2021ICMLIs space-time attention all you need for video understanding?Github starsTimeSformer
2021CVPRTaming transformers for high-resolution image synthesisGithub starsHigh resolution CNN+Transformer
2021ICCVEmerging properties in self-supervised vision transformersGithub starsDINO
SSL
2021ICCVVivit: A video vision transformerGithub stars---
2020ICLRAn image is worth 16x16 words: Transformers for image recognition at scaleGithub starsViT

关于 About

Paper Survey for Transformer-based SLAM
awesomedeep-learningdepth-estimationpose-estimationslamtransformer

语言 Languages

提交活跃度 Commit Activity

代码提交热力图
过去 52 周的开发活跃度
172
Total Commits
峰值: 30次/周
Less
More

核心贡献者 Contributors