Deployment Roadmap
Phase 0 — Freeze environment and inputs
- Complete the software record in Jetson Platform and Feasibility
- Configure active cooling and 25 W mode
- Lock the camera mode, focus, and exposure
- Complete calibration below 0.5 px mean reprojection error
- Record development and untouched evaluation sweeps
Gate: the final camera mode and Jetson software stack are reproducible.
Phase 1 — DA3 resource feasibility
- Install the basic official DA3 environment
- Download DA3 Small
- Run the 2-frame warm-up
- Apply the primary 4×336 five-window test
- Apply the 3×280 fallback only if required
- Lock and record the selected configuration
Gate: five windows complete below 7.0 GB, or the project records a feasibility stop.
Phase 2 — Output and coordinate validation
- Validate every output in DA3 Model and Outputs
- Visualize processed RGB, depth, and confidence
- Verify world-to-camera convention and matrix inversion
- Compare predicted and calibrated intrinsics diagnostically
- Save a complete diagnostic window
Gate: one window has valid outputs, coherent cameras, and verified coordinates.
Phase 3 — Single-window point cloud
- Apply the 40th-percentile confidence filter
- Back-project with predicted intrinsics
- Transform points with inverted extrinsics
- Attach processed-image colors
- Normalize first-window median retained depth to 1.0
- Downsample at 0.01 internal units
- Display, export, and reload the PLY
Gate: single-window geometry has correct color and coordinate conventions.
Phase 4 — Two-window alignment
- Preserve exact shared frames and processed grids
- Construct shared-frame 3D correspondences
- Implement RANSAC Sim(3)
- Refine with Umeyama alignment
- Apply every gate in Window Alignment and Mapping
- Transform points and cameras into global coordinates
- Append non-overlap geometry only
- Verify rejected windows change no global state
Gate: two windows align without visible duplication.
Phase 5 — Recorded bounded reconstruction
- Process the untouched sweep using the locked window
- Consolidate every five accepted windows
- Enforce 30-window and 1,000,000-point limits
- Refresh the viewer after every accepted window
- Export relative PLY and trajectory
- Confirm rejected windows contributed no points or poses
Gate: the recorded sequence produces a coherent bounded map.
Phase 6 — Live integration
- Implement the bounded stages in System Architecture
- Keep one active inference and at most one pending window
- Replace only stale non-overlap candidates
- Keep capture and preview responsive
- Update the viewer after accepted windows
- Produce at least one accepted update every five seconds
- Save all required runtime artifacts
Gate: the five-minute live performance protocol passes.
Phase 7 — Scale and final evaluation
- Apply the independent 1.0 m scale reference
- Save raw relative and metric-scaled outputs
- Measure doorway width and height
- Complete every required protocol
- Complete the 15-minute sustained run
- Populate Results Templates
Optional optimization
- Test ray pose once against the baseline
- Keep ray pose only if it fits memory/update limits and improves alignment
- Investigate TensorRT, C++/GGML, or reduced official streaming only after completion
Do not make optional optimization a graduation dependency.
Deliverables
- Reproducible environment and camera records
- Source implementation and frozen configuration
- Development and evaluation run logs
- Raw relative and metric-scaled PLY clouds
- Raw relative and metric-scaled trajectories
- Alignment and performance metrics
- Annotated screenshots and final demonstration
- Completed result tables and interpretation