Views
No views yet
script.py versions that scored worse than the current
best (0.4584) but did NOT completely fail. Kept here in case you want to
re-deploy them, A/B test variants, or build on them.s23-model/) so files here
will NEVER get pushed to the leaderboard accidentally.| File | HF commit | Leaderboard score | Notes |
|---|---|---|---|
script_colmap_junction.py + postprocess_v2_colmap_junction.py | 6cf3fbd | hss=0.4443, corner=0.5003, edge=0.4101 | COLMAP multi-view plane refinement + junction constraints + snap class 3 + safer snap_horizontal + loosened hybrid_merge. All bundled. -0.014 vs baseline. |
script_pipeline_ensemble.py | 3118106 | hss=0.4406, corner=0.4856, edge=0.4122 | Learned + classical (sklearn_submission) pipeline ensemble with heavy features disabled to avoid timeout. -0.018 vs baseline. The "lite ensemble" variant. |
script_ensemble_full_TIMEOUT.py | 9cbefb9 | TIMEOUT (0.0) | Full classical-pipeline ensemble — predict_wireframe_sklearn called with all default features ON (USE_TRACK_ENSEMBLE etc). Re-runs DLT triangulation that hybrid_merge already does. Hit the 2h budget. Worst possible outcome (all zeros). |
script_tta_unoptimized_TIMEOUT.py | 8e33a89 | TIMEOUT (0.0) | First TTA attempt with K=3 passes, but each pass re-ran build_compact_scene (the expensive multi-view label voting). 3x the slowest step → timeout. The fix is now in production: cache fusion, vary only priority sampling. |
script_degraded_~0.32.py | (synthetic, not on HF) | ~0.30–0.35 estimated | Deliberately degraded variant. Starts from the 0.4584 baseline (commit 56f1ec6) and: (1) raises CONF_THRESH 0.4→0.7 (drops mid-confidence segments), (2) disables snap_to_point_cloud + snap_horizontal (raw model coords without semantic refinement), (3) disables classical track ensemble (no predict_wireframe_tracks + hybrid_merge). The model's overconfident skeleton without any post-processing. |
script_colmap_junction.py — the individual fixes inside (snap class 3,
safer snap_horizontal) might help in isolation; the bundle is what hurt.
The COLMAP refine code is fully integrated and could be tuned (try
min_quality=0.7, max_displacement=0.2 instead of the loose defaults
that snapped vertices to wall planes).script_pipeline_ensemble.py — the predict_wireframe_sklearn plumbing
is wired up. If you ever want to do a SELECTIVE ensemble (e.g., only use
classical-pipeline output for samples where learned model has low
confidence), this is the starting point.1# Example: deploy the COLMAP-junction variant to the HF repo
2cp 1/script_colmap_junction.py s23-model/script.py
3cp 1/postprocess_v2_colmap_junction.py s23-model/s23dr_2026_example/postprocess_v2.py
4cd s23-model && git add . && git commit -m "Try variant X" && git pushscript.py at commit 857514e: TTA optimized (3-pass with cached fusion).
Awaiting leaderboard score.| Commit | Score | Description |
|---|---|---|
9deb9e1 | hss=0.4583 | Hybrid geometric ensemble (pre-experiment baseline) |
61a6827 | hss=0.4583 | Pure revert to 9deb9e1 |
56f1ec6 | hss=0.4584 | CONF_THRESH 0.5→0.4 (current production) |