Expected the same photo collection, run through the same pipeline, should produce comparable reconstruction quality.
Actual: Across multiple sessions on the same subject:
One run produced a mostly-clean reconstruction with the top of the head missing.
The same photo collection, copied to a second machine, produced a reconstruction missing roughly half the head.
A later run collapsed entirely into a flat 2D plane rather than a 3D surface.
Switching image matching from the default to Exhaustive improved results somewhat but did not fully resolve missing geometry or surface roughness.
Redoing the capture protocol (continuous coverage including crown and chin, not just a single horizontal sweep) reduced but did not eliminate holes.
What was tried:
Byte-level verification that "identical" photo collections across two machines were actually identical: file count (dir /b | find /c /v "") and MD5 hash comparison (certutil -hashfile) on every file. Confirmed identical; ruled out a data-transfer issue.
Version parity check: confirmed both machines were running the same Meshroom 2025.1.0 via where Meshroom.exe and file Properties.
Matching method: switched from the default to Exhaustive/pairwise.
Manual initial pair: set StructureFromMotion's Initial Pair A/B manually in standalone Meshroom (two images ~45-90° apart, both landmark-rich, not near-duplicate frames), which noticeably improved stability, this was the single most effective fix found, but is only available in standalone Meshroom (see Problem 3).
DepthMapFilter thresholds: relaxed Min Consistent Cameras (to 2) and Min Consistent Cameras Bad Similarity (to 3), per AliceVision's own documented guidance for exactly this symptom.
Capture protocol: moved from a single flat sweep to a full rotation with dedicated up/down passes; still produced holes, traced to insufficient overlap specifically at the transitions between motions rather than total coverage.
Is the run-to-run non-determinism in feature matching/initial-pair selection something Open-LIFU has characterized or mitigated internally? Is there a recommended/validated capture protocol (camera path, overlap percentage, image count) beyond what's in the 2024 Spotlight/Protocol writeup?
Expected the same photo collection, run through the same pipeline, should produce comparable reconstruction quality.
Actual: Across multiple sessions on the same subject:
One run produced a mostly-clean reconstruction with the top of the head missing.
The same photo collection, copied to a second machine, produced a reconstruction missing roughly half the head.
A later run collapsed entirely into a flat 2D plane rather than a 3D surface.
Switching image matching from the default to Exhaustive improved results somewhat but did not fully resolve missing geometry or surface roughness.
Redoing the capture protocol (continuous coverage including crown and chin, not just a single horizontal sweep) reduced but did not eliminate holes.
What was tried:
Byte-level verification that "identical" photo collections across two machines were actually identical: file count (dir /b | find /c /v "") and MD5 hash comparison (certutil -hashfile) on every file. Confirmed identical; ruled out a data-transfer issue.
Version parity check: confirmed both machines were running the same Meshroom 2025.1.0 via where Meshroom.exe and file Properties.
Matching method: switched from the default to Exhaustive/pairwise.
Manual initial pair: set StructureFromMotion's Initial Pair A/B manually in standalone Meshroom (two images ~45-90° apart, both landmark-rich, not near-duplicate frames), which noticeably improved stability, this was the single most effective fix found, but is only available in standalone Meshroom (see Problem 3).
DepthMapFilter thresholds: relaxed Min Consistent Cameras (to 2) and Min Consistent Cameras Bad Similarity (to 3), per AliceVision's own documented guidance for exactly this symptom.
Capture protocol: moved from a single flat sweep to a full rotation with dedicated up/down passes; still produced holes, traced to insufficient overlap specifically at the transitions between motions rather than total coverage.
Is the run-to-run non-determinism in feature matching/initial-pair selection something Open-LIFU has characterized or mitigated internally? Is there a recommended/validated capture protocol (camera path, overlap percentage, image count) beyond what's in the 2024 Spotlight/Protocol writeup?