Skip to content

Track upstream ODM changes: pixi build (merged), OpenSfM v1.1 and OpenMVS 2.4 quality PRs (open) #88

Open

Description

@muratmaga

Upstream ODM is landing a set of changes that may affect reconstruction quality and how ODM is installed. This issue tracks their progress. No local testing until the quality PRs are actually merged and finalized — this is a watch-and-wait item.

PRs being tracked

1. Build system → pixi — OpenDroneMap/ODM#2037 — ✅ MERGED (2026-08-21)

  • Moves the entire ODM build to pixi; cold build drops from ~40 min to ~7–30 min, works on Linux/Windows/Mac.
  • No intended quality change (dependency versions pinned to match the old build), but "an expert's view on output differences" is an open checkbox upstream, and the Windows installer is explicitly listed as not yet tested.
  • Relevance to us: the windows-only work drives a natively installed NodeODM/ODM on Windows, so our Windows install documentation will likely need revisiting when the next ODM release ships with the pixi-based installer.

2. Reconstruction quality: OpenSfM v1.1 — OpenDroneMap/ODM#2071 — 🟡 OPEN

  • Re-ports ODM's customizations onto upstream Mapillary OpenSfM v1 — several years of upstream SfM development over ODM's ~2021-era frozen fork.
  • Branch commits directly relevant to our video-derived frame sets: reworked rolling-shutter correction, "keep best SIFT features only", track-manager/bundle-adjustment updates, two-view reconstruction fixes.
  • Not relevant to us: the per-axis GPS accuracy plumbing (our turntable/video captures have no GPS). The planar SfM algorithm is dropped (mapped to incremental); we don't use it.

3. Dense quality: OpenMVS 2.2.0 → 2.4.0 — OpenDroneMap/ODM#2074 — 🟡 OPEN

  • Brings OpenMVS v2.3.0 + v2.4.0 (Jan 2026, ROI estimation). ODM pins old default behavior, so out-of-the-box output should be similar; the interesting part is four newly exposed knobs:
    • --pc-fusion-filter (merge / fuse / dense-fuse) — dense-fuse applies a stricter multi-view consensus before creating a point; plausible lever against floaters/noise in masked close-range specimen reconstructions.
    • --pc-fusion-depth-diff (relative depth tolerance, default 0.01) — meaningful at specimen scale.
    • --pc-fusion-reprojection, --pc-crop-to-roi.
  • Note: the #2074 branch includes the #2071 OpenSfM changes, so once both are finalized a single build covers both.

Caveats

  • Neither quality PR posts benchmark numbers — "improvement" is asserted upstream, not demonstrated. A before/after on our own data is the only answer that counts.
  • Nothing reaches the opendronemap/nodeodm Docker images or the Windows bundles our ODM module uses until these are merged and released.

Plan (after merge + release)

  1. Build ODM once from the released version (pixi run build).
  2. Run 2–3 reference masked sets we've already reconstructed well — one photo turntable set, one video-derived set — through pixi run odm with identical parameters on old vs. new.
  3. Compare OpenSfM stats (tie points, reprojection error, cameras registered) and cloud-to-mesh distance of the dense output; visually check texture.
  4. Extra arms on the new build: --pc-fusion-filter dense-fuse, and --rolling-shutter on the video set (our ODM module's parameter grid doesn't currently expose --rolling-shutter — candidate addition if it helps).
  5. If results are convincing, report back upstream (they are explicitly missing an outside expert's view on output differences) and update the ODM module's parameter grid / Windows install docs as needed.

Watch list

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions