Results Templates

Resource feasibility

FramesResolutionRuns completedPeak memoryInference p50/p95Decision
4336
3280

Record the fallback row only when the primary fails.

Camera calibration

Images acceptedResolutionMean reprojection errorMaximum errorFocus/exposure locked

Single-window outputs

WindowValid depthConfidence p40Retained pointsPredicted fx/fyPLY reload

Window alignment

AttemptsAcceptedRejectedSuccess rateMedian correspondencesMedian inlier ratioMedian residual

For rejected attempts, include a separate count by rejection reason.

Doorway after scale alignment

DimensionReferenceReconstructedAbsolute errorPercentage error
Width
Height

Also report the raw internal length of the independent 1.0 m reference and the resulting scale factor.

Live performance

Window configInference p50/p95Alignment p50/p95Update interval p50/p95Peak memoryAvg/peak powerMax temp
Locked configuration

Sustained run

DurationAccepted windowsRejected windowsFinal pointsPeak memoryMax p95 update intervalMax tempPass
15 min

Required figures

All report evidence must be generated by this project. External demonstrations in Final Demo Target and Reference Systems are presentation references only and must not be used as result figures.

FigureRequired view or comparisonRequired annotations and evidence
Processed RGB, relative depth, and confidenceThe same pixels from one accepted window, shown side by sideWindow ID, frame ID, processed resolution, confidence threshold, and shared color scale where applicable
Single-window cloud with predicted camerasOne oblique view that makes room structure and all window cameras visibleWindow ID, retained-point count, internal normalization, and coordinate axes
Two-window overlap before and after Sim(3)Matched viewpoints with identical crop, camera, point size, and colorsSource/target IDs, shared frames, correspondences, inlier ratio, residual, and estimated scale
Final bounded colored cloud and trajectoryOne overview plus one close structural view of the doorway, desk, and wall cornerAccepted/rejected windows, final point count, bound events, and relative or metric units
Independent scale referenceSource photograph and reconstructed endpointsPhysical 1.0 m length, selected endpoints, raw internal length, and resulting scale factor
Doorway width and heightSource photograph and matched reconstruction viewReference and reconstructed endpoints, dimensions, absolute error, and percentage error
Registration outcomesAccepted and rejected counts grouped by reasonAll attempted windows remain in the denominator
Window-stage latencyDistribution or timeline for capture, inference, alignment, mapping, and displayp50, p95, units, run duration, and locked window configuration
Fifteen-minute telemetrySynchronized time series for memory, point count, power, temperature, and accepted-update intervalMap-bound events, rejected windows, gaps, throttling, and failure markers

Use project-generated screenshots or plots at readable resolution. Preserve the original artifact or log behind every figure, and record the run ID in the caption so the evidence can be reproduced.

Reporting conventions

  • State whether the primary or fallback window was used
  • Report raw relative and metric-scaled results separately
  • Use p50 and p95 for latency and update interval
  • Preserve rejected attempts in all denominators
  • Label manually selected endpoints
  • Report all map consolidation and point-cap events
  • Distinguish preview frame rate from reconstruction update rate

Interpretation

Address these questions directly:

  1. Did DA3 Small fit the target Jetson stack?
  2. Was geometry coherent inside individual windows?
  3. Were overlapping windows joined reliably?
  4. What metric error remained after one independent scale alignment?
  5. Was the chunked live update rate convincing and stable?

Do not equate low overlap residual with ground-truth pose accuracy. Do not describe the scaled map as survey-grade. State the restricted scene and motion conditions.