Monocular semantic scene completion (SSC) for quadruped robots remains underexplored in real crowded indoor environments, where human-scene occlusion disrupts static geometry and human occupancy predictions are often incomplete or spatially misplaced. We present CrowdOcc, an RGB-D dataset and monocular SSC framework for this setting. CrowdOcc contains 25.1K frames from 11 indoor scenes, with semantic occupancy annotations constructed through static dynamic decoupling. Our framework combines: (i) Normal Guided Scene Geometry Fusion (NGSGF) to complement depth-aware lifting with surface-normal cues for occlusion robust geometry; and (ii) Human-Centric Sparse Interaction (HCSI) to selectively model human-human and local human scene relations in 3D. Our method achieves state-of-the-art SSC performance on CrowdOcc's scene-disjoint test set, reaching 15.80 IoU, 11.40 mIoU, and 46.23 Human IoU, demonstrating generalization to unseen indoor scenes.
Figures & tables
Fig. 1: CrowdOcc Overview. Synchronized RGB-D, LiDAR, and IMU observations collected by a quadruped robot are used for the offline construction of semantic occupancy annotations. The resulting dataset provides RGB-D frames and semantic occupancy annotations for monocular SSC, where inference takes a single RGB image as input. The provided depth maps can optionally serve as auxiliary supervision during training. The red background denotes the sensing hardware, the blue background indicates the CrowdOcc dataset construction pipeline, while the turquoise dashed box shows the model prediction pipeline.
Dataset
View
Plat.
Crowd
Move.
Env.
Modality
Volume Size
Voxel (m)
Frames
Anno.
NYUv2 [ 15 ]
Indoor
C&D
144×240×240
0.02
1.4K
Occ-ScanNet [ 9 ]
Indoor
C
36×60×60
0.08
65.5K
EmbodiedOcc-ScanNet † [ 16 ]
Indoor
C
Variable
0.08
20.2K
SemanticKITTI [ 10 ]
Outdoor
C&L
32×256×256
0.20
9K
SSCBench-KITTI-360 [ 17 , 18 ]
Outdoor
C&L
32×256×256
0.20
12.9K
Occ3D-nuScenes [ 11 , 19 ]
Outdoor
C&L
16×200×200
0.40
40K
Table 2
Fig. 2: Pipeline of our CrowdOcc dataset generation. We adopt a dual-branch pipeline for human-free and human-present sequences. The human-free branch reconstructs a dense semantic static scene, while the human-present branch extracts human observations and aligns them with the static scene. During voxelization, the synchronized current-frame depth map is used to refine human occupancy labels, followed by uncertainty and validity masking to produce the final semantic occupancy annotations.
Fig. 3: Pipeline of our CrowdOcc framework. Given a monocular RGB image, MoGe-2 estimates depth and surface normals from the same image. We use Normal-Guided Scene Geometry Fusion to enhance multi-scale 2D scene features before D-FLoSP lifting, and Human-Centric Sparse Interaction to selectively refine human-related 3D features. The enhanced representation is then decoded for semantic scene completion.
Method
Venue
Floor (28.97%)
Wall (35.19%)
Ceiling (13.93%)
Human (8.27%)
Door (6.53%)
Glass (6.30%)
Hydrant (0.55%)
Other (0.26%)
Precision ↑
Recall ↑
IoU ↑
mIoU ↑
SplatSSC [ 29 ]
AAAI 2026
15.74
5.36
0.84
21.11
0.50
3.00
0.34
0.13
19.61
15.92
9.63
5.88
GPOcc [ 30 ]
CVPR 2026
16.99
6.53
0.93
32.51
1.80
2.42
1.67
0.67
20.21
18.18
10.58
7.94
EmbodiedOcc [ 16 ]
ICCV 2025
19.54
6.76
1.40
30.89
1.10
2.13
2.41
0.10
19.39
21.60
11.38
8.04
MonoScene [ 8 ]
CVPR 2022
19.17
9.53
1.52
43.09
1.57
4.36
1.69
0.00
21.67
30.31
14.47
10.12
NDC-Scene [ 24 ]
ICCV 2023
20.66
9.68
2.09
44.12
3.34
1.22
1.46
1.21
20.15
34.00
14.48
10.47
ISO [ 9 ]
ECCV 2024
20.02
9.88
0.97
43.79
2.69
4.41
2.87
1.10
21.25
31.32
14.49
10.72
Table 5
Fig. 4: Qualitative comparison of CrowdOcc with other monocular SSC methods.
College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China · Robotics Institute, Carnegie Mellon University, Pittsburgh, PA, USA