fix: face traces split at scene cuts — even same person, different cut

scripts/store_traced_faces.py

@@ -62,8 +62,17 @@ def run_face_tracker(face_json_path: str, traced_json_path: str) -> str:
             }
 
     print(f"[TRACE] Processing {len(face_data.get('frames', {}))} frames")
-
+    
-    face_data = track_faces(face_data, use_embedding=True)
+    # Load cut boundaries from cut.json (same directory as face.json)
+    cut_boundaries = None
+    cuts_path = face_json_path.replace("_traced.json", ".cut.json").replace(".face.json", ".cut.json")
+    if os.path.exists(cuts_path):
+        with open(cuts_path) as f:
+            cuts = json.load(f)
+        cut_boundaries = {s["start_frame"] for s in cuts.get("scenes", []) if s["start_frame"] > 0}
+        print(f"[TRACE] Loaded {len(cut_boundaries)} cut boundaries")
+    
+    face_data = track_faces(face_data, use_embedding=True, cut_boundaries=cut_boundaries)
     metadata = face_data.get("metadata", {})
     metadata["tracking_method"] = "iou_embedding"
     metadata["tracked_at"] = datetime.now().isoformat()

scripts/utils/face_tracker.py

@@ -28,7 +28,7 @@ Output:
 import json
 import argparse
 import numpy as np
-from typing import Dict, List
+from typing import Dict, List, Set
 from collections import defaultdict
 
 
@@ -112,6 +112,9 @@ def match_faces(
     distance_threshold: float = 100.0,
     use_embedding: bool = True,
     frame_gap: int = 1,
+    cut_boundaries: Set[int] = None,
+    prev_frame: int = None,
+    curr_frame: int = None,
 ) -> Dict[int, int]:
     """
     Match current frame faces to previous frame faces
@@ -124,6 +127,9 @@ def match_faces(
         distance_threshold: Maximum bbox center distance for matching
         use_embedding: Whether to use embedding similarity
         frame_gap: Number of frames between current and previous (1=adjacent)
+        cut_boundaries: Set of frame numbers where scene cuts occur
+        prev_frame: Previous frame number (for cut detection)
+        curr_frame: Current frame number (for cut detection)
     
     Returns:
         Dict mapping current_face_index -> previous_face_index (or -1 if new)
@@ -131,6 +137,12 @@ def match_faces(
     if not previous_faces:
         return {i: -1 for i in range(len(current_faces))}
     
+    # If a scene cut exists between prev and current frame, force all new traces
+    if cut_boundaries and prev_frame is not None and curr_frame is not None:
+        for cf in cut_boundaries:
+            if prev_frame < cf <= curr_frame:
+                return {i: -1 for i in range(len(current_faces))}
+    
     matches = {}
     used_prev = set()
     
@@ -193,8 +205,6 @@ def match_faces(
             if prev_at_edge and not curr_at_edge and similarity < 0.8:
                 continue
             
-            if iou > iou_threshold and similarity > similarity_threshold:
-            
             if iou > iou_threshold and similarity > similarity_threshold:
                 score = iou + similarity
             elif iou > 0.5 and similarity > 0.65:
@@ -226,6 +236,7 @@ def track_faces(
     similarity_threshold: float = 0.7,
     distance_threshold: float = 100.0,
     use_embedding: bool = True,
+    cut_boundaries: Set[int] = None,
 ) -> Dict:
     """
     Track faces across all frames
@@ -254,6 +265,7 @@ def track_faces(
     prev_faces = []
     prev_trace_ids = []
     prev_frame_num = None
+    prev_face_frame = None  # last frame number that had actual faces
     
     print(f"\nTracking faces across {len(sorted_frames)} frames...")
     print(f"Parameters: iou={iou_threshold}, similarity={similarity_threshold}, distance={distance_threshold}")
@@ -279,6 +291,9 @@ def track_faces(
             distance_threshold,
             use_embedding,
             frame_gap,
+            cut_boundaries,
+            prev_face_frame,
+            frame_num,
         )
         
         trace_ids = []
@@ -307,6 +322,7 @@ def track_faces(
         
         prev_faces = faces
         prev_trace_ids = trace_ids
+        prev_face_frame = frame_num
         
         if frame_num % 100 == 0:
             print(f"  Frame {frame_num}: {len(faces)} faces, {len(set(trace_ids))} active traces")
@@ -455,9 +471,18 @@ def main():
     parser.add_argument("--similarity-threshold", type=float, default=0.7, help="Embedding similarity threshold")
     parser.add_argument("--distance-threshold", type=float, default=100.0, help="Distance threshold")
     parser.add_argument("--no-embedding", action="store_true", help="Disable embedding matching")
+    parser.add_argument("--cuts-json", help="Path to cut.json for scene-cut-aware tracking")
     parser.add_argument("--analyze-only", action="store_true", help="Only analyze, don't output")
     args = parser.parse_args()
     
+    # Load cut boundaries if provided
+    cut_boundaries = None
+    if args.cuts_json:
+        with open(args.cuts_json) as f:
+            cuts = json.load(f)
+        cut_boundaries = {s["start_frame"] for s in cuts.get("scenes", []) if s["start_frame"] > 0}
+        print(f"  Cut boundaries loaded: {len(cut_boundaries)} cuts")
+    
     print("=" * 60)
     print("Face Tracker")
     print("=" * 60)
@@ -474,6 +499,7 @@ def main():
         similarity_threshold=args.similarity_threshold,
         distance_threshold=args.distance_threshold,
         use_embedding=not args.no_embedding,
+        cut_boundaries=cut_boundaries,
     )
     
     analyze_traces(face_data)