Research clip
A 3D scene graph built live from one RGB camera, ScanNet scene 0000
Ten seconds of screen recording at sixteen times real time. One image stream goes in, and a layered map of a whole apartment comes out, with no depth sensor and no LiDAR anywhere in the loop.
The whole pipeline on ScanNet scene 0000. M2H-MX predicts depth and semantics from the image stream, Mono-Hydra++ consumes those predictions inside the visual-inertial odometry and the mapping backend, and the hierarchical scene graph is built in real time. Playback is 16x.
What you are seeing
The input is one image stream
Three views run down the right of the frame: the raw RGB frame from ScanNet scene 0000, the semantic labels M2H-MX predicts from it, and the depth M2H-MX predicts from the same frame. Those predictions are everything the map is built from. No depth camera and no LiDAR appear anywhere in this recording.
The map is a graph, not a point cloud
The structure on the left hangs from a single building node. A room node sits under it, a layer of place nodes under that marking the free space the camera has moved through, and below those the object boxes, labelled bed, cabinet, desk and table, standing on the reconstructed RGB mesh. Every layer is attached to the one above it, so a query can go from a room to the objects inside it without searching the mesh.
It is built while the camera moves
Mono-Hydra++ takes the two predictions into its visual-inertial odometry and its mapping backend as they arrive, so the graph in the recording is the state of the map at that instant rather than a reconstruction assembled afterwards. Playback is sixteen times real time.
Recording details
- System
- M2H-MX front end, Mono-Hydra++ backend
- Input
- Monocular RGB, ScanNet scene 0000
- Sensors
- No depth sensor, no LiDAR
- Playback
- 16x real time
- Audio
- None. Screen recording
- Length
- 0:10 (10 seconds)
- Frame
- 1600 x 686 pixels
- File
- MP4, H.264, 1.7 MB
- Published
- August 2026
The work behind it
Better per-frame depth and semantics measurably improve the map and the trajectory, not just the benchmark score.
- 6.91 cmaverage ATE on selected ScanNet sequencesMono-Hydra stack with M2H-MX, monocular. Down from 17.59 cm for monocular Go-SLAM, a 60.7 % reduction. RGB-D baselines in the same table sit at 7.02 to 7.15 cm.
- The paper, in full Accepted, ICRA 2026 SRRA Workshop
- M2H-MX how the system is put together
- Preprint on arXiv
- Code on GitHub
Another clip
Watch next
The same idea on a flying platform
One loop of a real building, mapped from the air.