AIM-SLAM: Dense Monocular SLAM via Adaptive and Informative Multi-View Keyframe Prioritization with Foundation Model
Authors: Jinwoo Jeon, Dong-Uk Seo, Eungchang Mason Lee, Hyun Myung
Organizations: The School of Electrical Engineering, KAIST (Korea Advanced Institute of Science and Technology), Daejeon, 34141, Republic of Korea · KAIST InnoCORE LLM, KAIST, Daejeon, 34141, Republic of Korea
Recent advances in geometric foundation models have emerged as a promising alternative for addressing the challenge of dense reconstruction in monocular visual simultaneous localization and mapping (SLAM). Although geometric foundation models enable SLAM to leverage variable input views, the previous methods remain confined to two-view pairs or fixed-length inputs without sufficient deliberation of geometric context for view selection. To tackle this problem, we propose AIM-SLAM, a dense monocular SLAM framework that exploits an adaptive and informative multi-view keyframe prioritization with dense pointmap predictions from visual geometry grounded transformer (VGGT). Specifically, we introduce the selective information- and geometric-aware multi-view adaptation (SIGMA) module, which employs voxel overlap and information gain to retrieve a candidate set of keyframes and adaptively determine its size. Furthermore, we formulate a joint multi-view Sim(3) optimization that enforces consistent alignment across selected views, substantially improving pose estimation accuracy. The effectiveness of AIM-SLAM is demonstrated on real-world datasets, where it achieves state-of-the-art pose estimation performance and accurate dense reconstruction results. Our system supports ROS integration, with code is available at https://aimslam.github.io/.
Figures & tables
Fig. 1: Comparison among MASt3R-SLAM [ 8 ] , VGGT-SLAM [ 9 ] , and the proposed AIM-SLAM. MASt3R-SLAM relies on a fixed two-view input, while VGGT-SLAM processes a fixed chunk of consecutive keyframes. In contrast , AIM-SLAM adaptively prioritizes a variable number of keyframes with high viewpoint overlap and information gain. By jointly optimizing these multi-view inputs in Sim(3) space, AIM-SLAM achieves accurate and globally consistent dense reconstruction.
Fig. 2: Overall architecture of AIM-SLAM. The frontend consists of (a) multi-view prioritization method via the proposed SIGMA module, followed by VGGT-based dense pointmap inference, and (b) joint multi-view Sim(3) optimization to mitigate short- and mid-term drift. The backend loop closure module performs global pose-graph optimization to ensure global consistency.
Fig. 3: Block diagram of the proposed SIGMA module, which consists of three stages: (i) geometry-based subset initialization via voxel overlap, (ii) information-driven re-ranking based on covariance reduction, and (iii) adaptive activation regulated by a stability test. After each multi-view optimization, updated poses and confidences recurrently trigger the re-ranking process.
Fig. 4: Example of a voxel-indexed keyframe map for computing view overlap. Each voxel stores the IDs of keyframes that observe it. Using the last keyframe as the query, the system counts shared voxels and selects the top-N overlapping keyframes to initialize the multi-view subset.
Fig. 5: Effect of the SIGMA module on keyframe uncertainty reduction. Compared with the case without re-ranking (lower), incorporating information-driven re-ranking (upper) significantly decreases keyframe uncertainty, computed as the inverse of the fused point confidence aggregated across observations during optimization. This shows that the SIGMA module effectively retrieves informative frames to refine the keyframe. Uncertainty is visualized in color, with higher values shown in warmer colors. Regions with pronounced differences are highlighted with red rectangles.
TUM RGB-D
Method
360
desk
desk2
floor
plant
room
rpy
teddy
xyz
Avg.
DeepV2D [ 59 ]
0.243
0.166
0.379
1.653
0.203
0.246
0.105
0.316
0.064
0.375
DeepFactors [ 60 ]
0.159
0.170
0.253
0.169
0.305
0.364
0.043
0.601
0.035
0.233
DROID-SLAM [ 11 ]
0.111
0.018
0.042
0.021
0.016
0.049
0.026
0.048
0.012
0.038
DPV-SLAM [ 61 ]
0.112
0.018
0.029
0.057
0.021
0.330
0.030
0.084
0.010
0.076
Calib.
MASt3R-SLAM [ 8 ]
0.049
0.016
0.024
0.025
0.020
0.061
0.027
0.041
0.009
0.030
TABLE I: Quantitative comparison of camera pose accuracy on the TUM RGB-D dataset, measured by the RMSE of absolute trajectory error (ATE, unit: m). We indicate the top three results as first , second , and third .
EuRoC
Method
V101
V102
V103
V201
V202
V203
MH01
MH02
MH03
MH04
MH05
Avg.
DeepV2D [ 59 ]
0.981
0.801
1.570
0.290
2.202
2.743
0.739
1.144
0.752
1.492
1.567
1.298
DeepFactors [ 60 ]
1.520
0.679
0.900
0.876
1.905
1.021
1.587
1.479
3.139
5.331
4.002
2.040
DROID-SLAM [ 11 ]
0.037
0.013
0.019
0.017
0.010
0.013
0.013
0.012
0.022
0.048
0.044
0.022
DPV-SLAM [ 61 ]
0.035
0.008
0.015
0.020
0.011
0.040
0.013
0.016
0.022
0.043
0.041
0.024
Calib.
MASt3R-SLAM [ 8 ]
0.040
0.019
0.027
0.020
0.025
0.043
0.023
0.017
0.057
0.113
0.067
0.041
TABLE II: Quantitative comparison of camera pose accuracy on the EuRoC dataset, measured by the RMSE of absolute trajectory error (ATE, unit: m). We indicate the top three results as first , second , and third . † denotes that the average is computed excluding divergent sequences.
Method
EuRoC
TUM RGB-D
Accuracy
Completion
Chamfer
Accuracy
Completion
Chamfer
VGGT-Long [ 10 ]
0.106
0.119
0.112
0.094
0.107
0.100
VGGT-SLAM [ 9 ]
0.246
0.216
0.231
0.109
0.127
0.118
MASt3R-SLAM [ 8 ]
0.108
0.072
0.090
0.097
0.113
0.105
AIM-SLAM (ours)
0.103
0.102
0.102
0.063
0.098
0.081
TABLE III: Quantitative comparison with state-of-the-art methods for dense reconstruction on the EuRoC dataset (left) and TUM RGB-D dataset (right). The best results are highlighted in bold , and the second-best results are underlined .
Fig. 6: Dense reconstruction results on the EuRoC dataset (top row) and TUM RGB-D dataset (bottom row). With its adaptive multi-view design, the proposed method achieves robust dense reconstructions across diverse environments.
Fig. 7: Ablation study of the proposed SIGMA module showing how the total number of views affects pose accuracy (ATE RMSE) for the 7 TUM RGB-D dataset and 7 EuRoC dataset. Recency-based denotes selecting the most recent consecutive keyframes to form the input subset W , as in prior methods [ 8 , 9 , 10 ] .
Method
ATE RMSE
EuRoC
TUM RGB-D
Ray only
0.138
0.061
Projection only
0.081
0.032
Hybrid (ray + projection)
0.072
0.031
TABLE IV: Ablation study of the proposed hybrid residual for joint multi-view Sim(3) pose optimization on the EuRoC dataset (left) and the TUM RGB-D dataset (right).
State Key Laboratory of Industrial Control Technology, College of Control Science and Engineering, Zhejiang University, Hangzhou, China, 310027. · Key Laboratory of Collaborative Sensing and Autonomous Unmanned Systems of Zhejiang Province, Hangzhou, China, 310027.