rayv9_learnt_baseline_snap

S23DR 2026 submission.

Pipeline

1. v1d heatmap model — 7-channel (gestalt+depth+ade) CNN predicts vertex heatmaps per view; spatial NMS across views → ~80 rays/scene
2. v9 voxel model — rays + SfM points → 6-channel 64×32×64 voxel volume → RayVoxelTransformerV7a → hybrid ray/dense NMS → world-space vertices
3. Learned baseline — point fusion (COLMAP + depth) → EdgeDepthSegmentsModel → wireframe (vertices + edges)
4. Snap+unmatched — midpoint-snap baseline verts to nearest v9 vert within 1m; append unmatched v9 verts

Validation results (100 scenes, hoho22k_2026_trainval)

Method	Mean HSS	Median HSS
Learned baseline	0.352	0.369
+ Snap to v9	0.411	0.453

Vertex F1@0.5=0.494, F1@1.0=0.685

Checkpoints

File	Description	Size
v1d_checkpoint.pt	v1d heatmap model (GroupNorm CNN)	12M
v9_checkpoint.pt	RayVoxelTransformerV7a	22M
baseline_checkpoint.pt	EdgeDepthSegmentsModel	102M