Pointmap: radius from KD-tree median nearest-neighbour distance × 0.6 (proper 2D-surface heuristic)

app.py

@@ -267,11 +267,17 @@ def _make_glb(image_pil_texture: Image.Image, pointmap_hwc: np.ndarray,
     centroid = pts.mean(axis=0).astype(np.float32) if len(pts) else np.zeros(3, np.float32)
     pts = pts - centroid
 
-    # Sphere radius ≈ median nearest-neighbour distance in this cloud.
-    # Quick proxy: use scene extent / cbrt(N) which approximates inter-point
-    # spacing for a uniformly sampled volume.
-    extent = float(np.ptp(pts, axis=0).max()) if len(pts) else 1.0
-    radius = max(extent / max(np.cbrt(len(pts)), 1.0) * 0.7, 1e-4)
+    # Sphere radius from MEASURED median nearest-neighbour distance.
+    # Why measured (vs. extent / cbrt(N)): our points sit on a ~2D surface
+    # (skin), not in a 3D volume — a kd-tree gives the right spacing directly.
+    # Factor 0.6 → adjacent balls overlap slightly, forming a solid surface
+    # without big gaps and without obvious overlap blobs.
+    if len(pts) >= 2:
+        from scipy.spatial import cKDTree
+        nn_dist = cKDTree(pts).query(pts, k=2)[0][:, 1]   # k=1 is self
+        radius = max(float(np.median(nn_dist)) * 0.6, 1e-4)
+    else:
+        radius = 0.005  # fallback for empty/tiny clouds
 
     # Icosphere template (subdivisions=1 → 42 verts, 80 faces — smoother balls).
     sphere = trimesh.creation.icosphere(subdivisions=1, radius=radius)