We propose SLAM++, a new SLAM system that operates at the level of 3D objects rather than low-level geometric primitives. Given a database of known 3D object models, our system simultaneously recognizes objects, estimates their 6-DoF poses, and builds an object-level map of the environment in real-time using an RGB-D camera. The resulting map is dramatically more compact and semantically meaningful than traditional point cloud or surfel maps — a room can be described by a handful of object instances with poses rather than millions of points. This representation also enables higher-level reasoning about scene structure, object interactions, and relocation.

Traditional SLAM systems build maps as collections of geometric primitives — points, lines, surfels, or dense volumetric grids. While these representations capture the geometry of the environment, they are semantically impoverished: a point cloud map does not "understand" that certain clusters of points form a chair, a table, or a monitor. This lack of semantic structure limits the utility of SLAM maps for high-level tasks such as manipulation planning, human-robot interaction, and scene understanding.

We argue that the right level of abstraction for indoor SLAM is objects. In structured indoor environments such as offices and living rooms, the scene is largely composed of repeated, known object categories (chairs, tables, monitors, keyboards). If the system can recognize these objects and estimate their poses, the map becomes a compact graph of object instances — orders of magnitude smaller than raw point clouds, yet capturing both geometry and semantics. This paradigm shift from "mapping everything" to "mapping what matters" also offers computational benefits.

KinectFusion demonstrated real-time dense 3D reconstruction using an RGB-D sensor with a truncated signed distance function (TSDF) volume. While impressive, the resulting maps are purely geometric and scale poorly to large environments. ORB-SLAM and other feature-based systems produce sparse maps that are efficient but lack surface detail and semantic information. Recent works on semantic mapping add object labels to existing maps via post-hoc segmentation, but do not leverage object models as first-class map elements.

SLAM++ operates in real-time on RGB-D data from a Kinect-like sensor. The system maintains a pose graph where nodes represent either camera keyframes or recognized object instances, and edges represent geometric constraints between them. For each incoming frame, the system performs three operations: (1) camera tracking via ICP alignment against the current map, (2) object recognition and pose estimation using a model database, and (3) graph optimization to jointly refine all camera and object poses. Newly detected objects are added as nodes; re-observed objects provide loop closure constraints.

Object recognition is performed by matching depth data against a database of pre-scanned 3D models. We use a multi-resolution approach: first, candidate object hypotheses are generated by matching local surface descriptors (FPFH features) between the current depth frame and the model database. Then, each hypothesis is refined via ICP alignment between the observed depth and the model, with the alignment quality serving as a verification score. Objects must achieve an ICP fitness above a threshold to be accepted. The 6-DoF pose from ICP provides the geometric constraint between the camera and the object in the pose graph.

A key benefit of object-level mapping is map compactness. A traditional dense map of an office scene might contain millions of surfels or a large TSDF volume. In contrast, SLAM++ represents the same scene as a graph with perhaps 20-30 object nodes, each storing only a model ID and a 6-DoF pose (7 parameters). This yields a compression ratio of several orders of magnitude. Moreover, the object-level map is immediately useful for tasks like robotic grasping — the system knows not just "there is geometry here" but "there is a mug at this pose."

We evaluate SLAM++ on real-time RGB-D sequences captured in office environments with a Kinect sensor. The object database contains pre-scanned models of common office objects including chairs, monitors, and keyboards. The system runs at approximately 20 Hz on a desktop GPU, including object recognition. In quantitative evaluation, SLAM++ achieves camera tracking accuracy comparable to KinectFusion while producing a map that is over 1000x more compact. Object pose estimation accuracy is within 2-3 cm translation and 5 degrees rotation for well-observed objects.

We demonstrate several unique capabilities enabled by object-level mapping. First, relocation: when the camera is lost, recognizing a previously mapped object instantly recovers the camera pose without requiring visual feature matching. Second, map reuse across sessions: the compact object-level map can be saved and reloaded, allowing immediate localization in a previously mapped environment. Third, semantic queries: the system can answer questions like "where is the nearest keyboard?" directly from the map.

SLAM++ demonstrates that operating at the level of objects is a viable and advantageous paradigm for real-time SLAM. By replacing millions of low-level primitives with a compact graph of recognized object instances, the system achieves comparable tracking accuracy with dramatically reduced map size and gains semantic understanding for free. The current limitation of requiring a pre-built object database could be addressed in future work by integrating online object learning or category-level pose estimation. We envision SLAM++ as a step toward truly intelligent spatial understanding.