WatsonTang98/s23dr-2026-somnathab3-test

S23DR 2026 — Submission v2 (Tuned Post-Processing)

Learned-model submission for the S23DR 2026 wireframe-estimation challenge.

What changed vs v1

Same checkpoint as the original submission (checkpoint.pt, 170k steps, 4096-pt), but with post-processing hyperparameters tuned via grid search on the local 1024-scene validation set:

Parameter	v1 (default)	v2 (tuned)	Effect
MERGE_THRESH (merge_end)	0.55	0.40	Less aggressive vertex merge → better recall
SNAP_RADIUS	0.50	0.70	Larger point-cloud snap radius → better precision

Local val results (1024 scenes, 4096-pt):

Version	HSS@0.5	vF1	eIoU
v1 — original submission (default params)	0.376	0.428	0.323
v2 — tuned post-processing	0.390	0.457	0.353
Δ	+0.014	+0.029	+0.030

Pipeline

raw multi-view sample
  -> point fusion          per-view depth unprojection + COLMAP, labeled
                           by ADE/Gestalt              (point_fusion.py)
  -> priority sampling     4096 pts: 3072 COLMAP + 1024 depth
                                                       (make_sampled_cache.py)
  -> Perceiver             hidden=256, 256 latents x 7 layers, 64 segment
                           queries; (midpoint, direction, length) + confidence
                                                       (model.py)
  -> postprocess           conf > 0.5 -> segments -> iterative vertex merge
                           (end=0.40) -> snap to point cloud (r=0.70)
                           -> horizontal snap
  -> {order_id, wf_vertices, wf_edges}                 -> submission.json

Run inference

python script.py

The competition harness provides params.json and runs script.py. The script downloads the dataset to /tmp/data, loads checkpoint.pt, iterates over validation + test splits, and writes submission.json.

Requirements: CUDA GPU, torch, huggingface_hub, datasets, numpy, opencv-python, scipy, tqdm.

Model details

Property	Value
Architecture	Perceiver IO — TokenTransformerSegments
Parameters	~8.85 M
Input	[B, 4096, 95] — xyz + normals + local PCA features
Output	[B, 64] segment predictions
Checkpoint	checkpoint.pt (170k steps, same as v1)

Post-processing sweep

Grid search results (200 scenes per config):

merge_end	snap_radius	HSS@0.5
0.4	0.7	0.379 ← chosen
0.4	0.5	0.376
0.5	0.7	0.372
0.5	0.5	0.371
0.6	0.7	0.368
0.5	0.3	0.365
0.6	0.5	0.364
0.7	0.5	0.359
0.6	0.3	0.359
0.7	0.3	0.354