Partition on a globally verified view graph

[kathirgounder]

Stacked on #1134 (the inline-frontend PR — its commit appears here until it merges; review the last commit only).

What

The hierarchical pipeline currently partitions the retrieval graph: METIS carves clusters along MegaLoc-similarity edges, many of which have no verifiable two-view geometry. This PR runs the frontend (correspondence generation + two-view estimation) once over the full retrieval graph, up front, and hands everything downstream — partitioning, per-cluster reconstruction, merging — the verified subgraph (edges where TwoViewResult.valid()).

1. scene_optimizer.verify_visibility_graph (module function, tested): runs the frontend over all retrieval edges and returns padded per-image keypoints + verified correspondence indices per surviving edge. Dispatches on pool size — multiple workers fan out via create_v_corr_idxs_futures (only lean per-chunk sub-dicts are gathered); a single worker runs inline in the main process (no scheduler↔worker comm surface for a multi-hour run to trip over). Every run_2view call warms the two-view cache, so per-cluster frontends afterwards are cache hits.
2. Always on when verifiable — no flag. Any cluster optimizer exposing correspondence_generator + two_view_estimator (ClusterMVO, ClusterVGGTWithFrontend, incl. Cacher-wrapped) gets the verified graph. Frontend-less optimizers (ClusterVGGT, AnySplat, FastVGGT — pure feedforward) have no two-view stage to verify with and keep the retrieval graph; that's capability detection, not an A/B switch.
3. Global correspondences reused per cluster. ClusterContext gains global_v_corr_idxs_dict / global_keypoints; ClusterVGGTWithFrontend builds each cluster's 2D tracks by subsetting them to the cluster's edges (eagerly, main process) instead of re-running the per-cluster frontend, which would recompute the identical per-edge result serially across overlapping clusters. Data-driven: no globals in the context (optimizer used standalone) → the traditional per-cluster frontend runs, unchanged.
4. runner.py: Dask comm timeouts 30s → 300s (connect + tcp). Separate bug fix that this PR makes load-bearing: bulk image loads and the in-process frontend stall the client event loop past the 30s defaults, closing comms mid-run (CommClosedError / "lost dependencies") on large single-node scenes.
5. Golden phototourism config (vggt_sift_frontend_megaloc_phototourism.yaml): retrieval now over-generates (num_matched 15→100, min_score 0.5→0.15) because verification supplies the precision; METIS sizes up (30/70); ColmapSIFT@8192 + PoseLib LO-RANSAC@2px + inlier support 30/0.15 — the peak verifier stack the paper numbers used. Calibration/BA-prior keys are deliberately untouched (next PRs).

Tests

• New tests/test_verify_visibility_graph.py: the inline (1-worker) and fan-out (2-worker) arms produce identical keypoints + verified graphs; only estimator-validated edges survive.
• Full affected suite (incl. De-nest the cluster frontend (no nested Dask clients); memory-bounded two-view helpers #1134's tests): 23 passed.

Review focus

• scene_optimizer.py: the capability check (getattr for correspondence_generator/two_view_estimator — the Cacher forwards public attrs) and the to_context plumbing.
• cluster_vggt_with_frontend.py: the reuse branch skips _build_frontend_output_graphs (two-view viz/metrics) — per-edge results are identical to the per-cluster frontend, so nothing else changes downstream.
• Behavior change sign-off: classic Multiview configs now also partition on their verified graph (same frontend, run once globally; per-cluster passes cache-hit).

Context

Series landing the ECCV'26 SFM-DL hierarchical-SfM pipeline (research branch vggt-verified-pipeline, tag eccv2026-sfmdl-research) into mapping-vggt. Previous: #1134. Next: calibration arbitration (EXIF passthrough with model-focal fallback).

🤖 Generated with Claude Code