Mono-Hydra builds a Hydra-style 3D scene graph from a single camera plus an IMU, with no depth sensor. Learned networks supply depth and semantics from the RGB image, and a robocentric, square-root-information visual-inertial odometry (R-VIO2) provides the trajectory. The motivation is payload: RGB-D and LiDAR are heavy for small UAVs. The public ROS 2 repository (Mono-Hydra++) describes the current stack: R-VIO2 odometry, a pose graph with Kimera-style loop-closure proposals, a multi-task network (M2H) that publishes metric depth and semantic labels, and the official MIT SPARK Hydra backend for mesh, objects, places, rooms and the scene graph.
Findings
- The arXiv abstract reports sub-20 cm error at 15 fps on a laptop GPU (Nvidia 3080), mainly indoors.
- The README lists launch files for an indoor campus sequence, 7-Scenes, uHumans2 and ScanNet, and ablation switches for sparse depth factors, semantic masking and IMU-camera time alignment.
Limits
- Seedling note: I read only the arXiv abstract and the repository README, so I cannot say how depth errors propagate into places and rooms.
- Learned monocular depth carries scale and domain errors that a depth sensor does not; the whole graph inherits them.
- Model weights are not in the public repository, per the README.
Relevance
A small drone in a team can afford to carry camera-plus-IMU scene graphs, so this line of work is relevant to aerial teams and to multi-robot perception where robots have unequal sensor suites. A monocular robot’s places and rooms will be noisier than a LiDAR robot’s, and fusion schemes that weight every robot equally would not know that.