Bavantha Udugama 0000-0002-1932-692X

Research clip

A drone flies one corridor loop and the 3D scene graph builds itself

Forty-two seconds at ten times real time, from the first frame to loop closure. A camera and an IMU go in; a mesh, objects, places and rooms come out, and the corridor circuit closes on itself.

Mono-Hydra + M2HITC building, 2nd floor

The scene graph forming over one loop of the ITC second floor, from the first frame to loop closure. Playback is 10x real time. Live RGB, predicted semantics, and predicted depth run down the right.

Key moments

  1. 0:00Title card, then the first stretch of corridor
  2. 0:08The five layers labelled: building, rooms, places, objects, mesh
  3. 0:16The first turn, and room nodes appear under the building node
  4. 0:24Three sides of the circuit mapped
  5. 0:36Loop closure, seen from above

What you are seeing

One loop of a real building

The recording follows a drone through a single circuit of the second floor of the ITC building at the University of Twente. The camera stream, the semantic labels M2H predicts from it, and the predicted depth run down the right of the frame throughout, and the map on the left is built from those predictions alone.

Five layers, built together

Eight seconds in, the recording labels the layers: the building node at the top, the room nodes under it, a layer of semantic places, the object boxes below those, and the 3D RGB mesh at the bottom. The mesh extends along each corridor as the drone passes down it, and room nodes appear under the building node as the circuit turns.

Loop closure

In the last seconds the view turns top down and the circuit meets itself. The two ends of the corridor join, and the graph is corrected into one consistent floor instead of a drifting chain of corridors.

Recording details

System
M2H front end, Mono-Hydra backend
Input
Monocular RGB and IMU, carried on the drone
Place
ITC building 2nd floor, University of Twente
Playback
10x real time
Audio
None. Screen recording
Length
0:42 (42 seconds)
Frame
1280 x 570 pixels
File
MP4, H.264, 4.5 MB
Published
August 2026

The work behind it

Four dense prediction tasks share one small model and make each other better, fast enough to sit inside a mapping loop.

More clips