Model
Depth Anything V2
Estimate a dense depth field from a single ordinary camera, with no stereo rig.
shipped
Monocular Depth
DepthSpatialCUDA
Overview
Depth Anything V2 predicts relative inverse-depth (disparity) per frame, larger = nearer. Overseer normalises it to [0,1], temporally median-fuses several frames to suppress monocular noise on the static scene, and back-projects it through a pinhole model to lift the frame into 3D. The depth field also drives the depth-locked tactical radar.
| Role | Monocular Depth |
|---|---|
| Expected latency | ~120–400 ms / frame (Large tier, GPU); cached for the static scene |
| Hardware requirements | NVIDIA GPU strongly recommended |
| Memory requirements | ~2–4 GB VRAM (tier dependent) |
Inputs & Outputs
Inputs
- RGB frame
- Working resolution
- Fusion frame count
Outputs
- Normalised disparity grid (float32)
- Point cloud (with RGB)
- Ground / background depth
Advantages & Limitations
Advantages
- No calibration or stereo needed
- Clean, temporally stable geometry
- Great zero-shot generalisation across scenes
Limitations
- Relative, not metric, without calibration
- Reflective / transparent surfaces are hard
- Heavier than detection, run periodically, not every frame
Example outputs
Depth Anything V2
A coloured 3D point cloud of a street
Depth Anything V2
A ground-plane depth field behind occluding objects
Depth Anything V2
Depth-locked contact placement on the radar
Configuration
Keys in config/default.yaml that govern this model.
| Key | Default |
|---|---|
spatial.enabled | true |
spatial.input_width | 640 |
spatial.fuse_frames | 3 |
spatial.fov_deg | 60 |
Benchmarks
Placeholder
Benchmark figures below are illustrative and awaiting a standardised harness. Treat them as placeholders.
| Benchmark | Dataset | Metric | Value | Hardware |
|---|---|---|---|---|
| δ<1.25 | NYU (ref) | accuracy | ~0.98 | reference |
| Scene depth | 640px | ms | ~240 | RTX 3060 |