Mono-Hydra++: Real-Time Monocular Scene Graph Construction with Multi-Task Learning for 3D Indoor Mapping

2 min read

Mono-Hydra++ is the follow-up to Mono-Hydra: still one RGB camera plus an IMU, but now a single multi-task network (M2H-MX, a frozen DINOv3 backbone with LoRA) predicts both depth and semantics. The predicted depth also feeds back into the VIO as sparse metric constraints, the semantic output masks dynamic regions out of motion estimation, and recent predictions are warped by the VIO poses before fusion to reduce flicker. The graph itself is built by the Mono-Hydra backend, which builds on Hydra.

Findings

  • On the Go-SLAM ScanNet subset, average trajectory error is 1.6% lower than the strongest RGB-D baseline in their comparison, using only RGB and IMU.
  • On calibrated 7-Scenes, average trajectory error improves by 29.8% over the strongest calibrated baseline.
  • The perception model alone, exported with TensorRT FP16, runs at 25.53 FPS on a Jetson Orin NX 16GB, at reduced input resolution and with some accuracy loss.

Limits

  • The authors say the main remaining problem is losing small objects during semantic fusion, which shows up in the graph as missing object nodes and wrong object-room links.
  • The Jetson number covers the network, not the whole mapping pipeline; the main timings use a desktop RTX 4080 Super.
  • Labels appear to be a fixed ScanNet-style set; I have not seen any open-vocabulary layer. Indoor only.
  • Seedling: abstract, intro and discussion only.

Relevance

Mono-Hydra++ argues for scene graphs on a small drone with no depth sensor, although nothing here flies yet. It also shows the cost of skipping depth: object nodes are the first thing to go. See 3d-scene-graphs and aerial swarms.