Point Cloud Indoor Dataset

Indoor robots -- home assistants, office delivery bots, cleaning robots, retail inventory systems, and facility management platforms -- must build and reason about 3D representations of their environments. Point clouds provide the richest 3D representation available: millions of measured 3D points that capture room geometry, furniture placement, object locations, and navigable space with centimeter-level precision. Unlike 2D images or even depth maps, point clouds represent the full 3D structure of a scene, enabling robots to plan paths around furniture, reach for objects on shelves, and understand spatial relationships that are ambiguous in 2D projections.

Training 3D perception models on point cloud data is fundamentally different from training on images. Point clouds are unordered, irregular, and variable in density -- a nearby wall produces dense points while a distant corner may have sparse coverage. Models must be invariant to point ordering (PointNet), capture local geometric patterns (PointNet++, KPConv), or operate on voxelized representations (MinkowskiNet, SparseConvNet). Each approach requires training data with per-point annotations that accurately reflect the 3D semantic structure of real indoor environments.

Existing indoor point cloud datasets like ScanNet and S3DIS have driven significant progress in 3D scene understanding research, but they are limited in scale and diversity. ScanNet contains 1,513 scans of primarily academic and residential rooms. S3DIS covers 6 areas of a single university building. Real-world indoor robots encounter far more diverse environments: homes with varied architectural styles, offices with different furniture systems, retail stores with diverse shelving, hospitals with specialized equipment, hotels, restaurants, gyms, libraries, and more. Claru's indoor point cloud dataset captures 500+ rooms across 8+ room types with dense semantic annotations.

Research from CVPR 2024 and 3DV 2024 demonstrates that 3D scene understanding models trained on diverse indoor environments achieve 25-35% better generalization to novel room types compared to models trained on single-building datasets, with the improvement driven by exposure to varied furniture styles, room scales, and architectural configurations that build robust feature representations.

Primary scanning uses a Leica BLK360 G2 terrestrial laser scanner (360-degree coverage, up to 360K points/second, range accuracy +/-4mm at 10m, integrated HDR panoramic camera for colorized point clouds). Each room receives 2-4 scan positions to ensure complete coverage with no shadow zones behind furniture. Individual scans are registered into a unified coordinate frame using ICP alignment with a target registration error below 5mm.

Supplementary scanning with handheld LiDAR (Ouster OS0-32, carried through the space) provides dynamic scan sequences for training SLAM algorithms. The handheld sequences capture the temporal progression of map building as a robot would experience it -- starting from an unknown room and progressively discovering the space through exploration. Intel RealSense D455 cameras co-captured with the handheld sequences provide aligned RGB-D frames for training multi-modal 3D perception systems.

The dataset spans 8+ indoor environment types: residential living spaces (apartments, houses, studios), commercial offices (open plan, private offices, conference rooms), retail spaces (stores, showrooms), hospitality (hotel rooms, lobbies, restaurants), healthcare (patient rooms, waiting areas, clinics), educational (classrooms, labs, libraries), fitness (gyms, studios), and industrial (workshops, server rooms). Each environment type includes 50+ rooms with varied layouts, furniture configurations, and architectural styles to ensure diversity within categories.

Environmental metadata per scan includes room type, approximate floor area, ceiling height, number of distinct furniture items, dominant materials (wood, carpet, tile, concrete), lighting type (natural, fluorescent, LED), and clutter level (minimal, moderate, dense). For residential spaces, architectural style (modern, traditional, industrial, Scandinavian) is documented. This metadata enables researchers to study how 3D perception performance varies with environment characteristics and to build room-type-aware models.

Stage one automated processing generates: colorized point clouds from the BLK360 scans (RGB color transferred from the integrated panoramic camera), floor plane detection and room boundary extraction, initial semantic segmentation using pre-trained Mask3D, and navigable space estimation from the floor plane with obstacle margins. Point cloud density is normalized to ensure consistent annotation quality across rooms scanned from different distances.

Stage two human annotation adds per-point semantic labels across 40+ indoor object categories following a taxonomy compatible with ScanNet but extended with additional classes for commercial and hospitality environments: furniture (tables, chairs, desks, beds, sofas, shelving), fixtures (sinks, toilets, light fixtures, HVAC vents), electronics (monitors, TVs, printers), architectural elements (walls, floors, ceilings, doors, windows, columns, stairs), and room-specific items (kitchen appliances, bathroom fixtures, retail displays, gym equipment). Instance segmentation separates individual objects within each class.

Stage three adds higher-level spatial annotations: room layout estimation (wall planes, floor-ceiling boundaries), functional zone delineation (work area, circulation path, storage zone, social area), navigable space mapping with clearance annotations (wheelchair-accessible paths, narrow passages, step hazards), and object affordance labels (sittable, openable, graspable, pushable). These spatial annotations go beyond per-point semantics to capture the functional structure that indoor robots need for task planning.

Stage four QA combines automated geometric checks with human review. Per-point label consistency is verified across overlapping scan regions -- the same chair must receive the same label from every scan position. Instance segmentation boundaries are checked against the 3D geometry (instance boundaries should align with geometric discontinuities). Navigable space annotations are verified against actual room accessibility. Overall targets: 95%+ per-point semantic accuracy, 90%+ instance segmentation mAP, and 97%+ navigable space accuracy.

Claru's collector network deploys terrestrial laser scanners and handheld LiDAR across diverse indoor environments -- from luxury apartments to budget hotels, co-working spaces to medical clinics, retail stores to fitness centers. This diversity ensures that 3D perception models trained on the data generalize across the full range of indoor environments that robots encounter in deployment, not just the academic and residential settings covered by existing public datasets.

Custom campaigns can target specific room types (residential only, or office focus), architectural styles, furniture density levels, or accessibility requirements (ADA-compliant space annotations). For teams building simulation environments, Claru can provide textured mesh reconstructions alongside point clouds. Turnaround from campaign specification to annotated delivery is typically 4-8 weeks.

Data is delivered as colorized point clouds (PLY, PCD, LAS) with per-point semantic and instance labels. Accompanying files include room layout parameters, navigable space maps, camera calibration for RGB-D sequences, and handheld LiDAR trajectories for SLAM benchmarking. Format conversion to voxelized representations, ScanNet-compatible formats, or custom schemas is available at no additional cost.