Switch to alternative vertex generation strategies since DGCNN cannot run

sklearn_submission.py

@@ -106,7 +106,8 @@ DGCNN_MAX_DIST_TO_CLOUD = 5.0
 #   t=0.3 ΔHSS=−0.0018   t=0.5 +0.0021   t=0.6 +0.0030
 #   t=0.7 +0.0039 (peak)  t=0.8 +0.0031
 # Clean signal: F1 stable (±0.0006), IoU +0.0065 at t=0.7.
-USE_DGCNN_EDGES = True
+# Since s23dr is missing, DGCNN is impossible to run. We must disable it so it doesn't crash or waste time.
+USE_DGCNN_EDGES = False
 # Ask the edge model for a wider candidate set, then apply our own
 # geometry gates below. This recovers medium-confidence true edges without
 # letting the classifier densify the graph unchecked.
@@ -126,7 +127,8 @@ DGCNN_EDGE_REPROJ_DILATE_PX = 6
 # Regressed: +2 vertices and +2 edges per sample but the new vertices are
 # mostly ghosts. Need PointNet Stage 2 (vertex refinement model) to make
 # this useful — that requires training on ~600k samples from the dataset.
-USE_WINNER_CANDIDATES = False
+# Use winner 3D candidates to improve vertex recall
+USE_WINNER_CANDIDATES = True
 WINNER_DEDUP_RADIUS = 0.5
 WINNER_MAX_DIST_TO_CLOUD = 8.0
 
@@ -255,7 +257,8 @@ ISOLATED_TRACK_MAX_DIST = 3.5
 # F1 + IoU both signed positive (rare among our killed experiments) but
 # HSS delta is in noise range. Code path kept behind the flag for future
 # tuning or for combination with other refinements.
-USE_TRACKS_AS_VERTICES = False
+# Use triangulation tracks to refine and augment vertices
+USE_TRACKS_AS_VERTICES = True
 TRACK_MIN_VIEWS = 3
 TRACK_MAX_REPROJ_PX = 2.0
 TRACK_REPLACE_RADIUS = 0.6