Add 5% confidence threshold, limit to top 5 candidates, add tracking logs

Detection:
- conf_threshold: 0.0 -> 0.05 (5% minimum to filter noise)
- max_candidates: 5 (keep only top 5 by confidence)

Tracking:
- Confirmed: tracks across ALL frames (for idx, frame in enumerate(frames))
- Added logging to show tracking progress and results

Logs now show:
- [detect_all_balls] Found N candidates (top 5, conf >= 5%)
- [_track_single_ball_candidate] Tracking Ball N across M frames...
- [_track_single_ball_candidate] Ball N done: X/M frames (Y%), max_vel, kick info

app.py

@@ -135,16 +135,17 @@ def detect_ball_center(
 def detect_all_balls(
     frame: Image.Image,
     model_filename: str = YOLO_DEFAULT_MODEL,
-    conf_threshold: float = 0.0,  # Same as original - accept all detections
+    conf_threshold: float = 0.05,  # Minimum 5% confidence to filter noise
     iou_threshold: float = YOLO_IOU_THRESHOLD,
-    max_detections: int = 10,  # Increased to catch more candidates
-    roi_x_min: float = 0.0,  # NO ROI filter - accept balls anywhere
-    roi_x_max: float = 1.0,
+    max_detections: int = 10,  # Get more from YOLO, then filter to top 5
+    max_candidates: int = 5,  # Return only top 5 by confidence
 ) -> list[dict]:
     """
     Detect all ball candidates in a frame.
     
-    NO ROI filtering - we show ALL detected balls and let the user/algorithm decide.
+    - Minimum 5% confidence to filter noise
+    - Returns top 5 candidates by confidence
+    - No ROI filtering - scoring happens later
     
     Returns list of dicts with keys:
         - id: int (candidate index)
@@ -219,14 +220,17 @@ def detect_all_balls(
     # Sort by confidence descending
     candidates.sort(key=lambda c: c["conf"], reverse=True)
     
-    # Re-assign IDs after sorting
+    # Keep only top N candidates
+    candidates = candidates[:max_candidates]
+    
+    # Re-assign IDs after sorting and filtering
     for i, c in enumerate(candidates):
         c["id"] = i
 
     # Debug logging
-    print(f"[detect_all_balls] Found {len(candidates)} ball candidates:")
+    print(f"[detect_all_balls] Found {len(candidates)} ball candidates (top {max_candidates}, conf >= {conf_threshold:.0%}):")
     for c in candidates:
-        print(f"  Ball {c['id']}: center={c['center']}, conf={c['conf']:.3f}, box={c['box']}")
+        print(f"  Ball {c['id']}: center={c['center']}, conf={c['conf']:.1%}, box={c['box']}")
 
     return candidates
 
@@ -788,7 +792,7 @@ def _track_single_ball_candidate(
     progress: gr.Progress | None = None,
 ) -> dict:
     """
-    Track a single ball candidate across all frames using YOLO.
+    Track a single ball candidate across ALL frames using YOLO.
     Uses proximity matching to follow the same ball.
     
     Returns dict with tracking results:
@@ -807,6 +811,8 @@ def _track_single_ball_candidate(
     frames = state.video_frames
     total = len(frames)
     
+    print(f"[_track_single_ball_candidate] Tracking Ball {candidate['id']} across {total} frames...")
+    
     # Initial position from candidate
     last_center = candidate["center"]
     max_distance_threshold = 100  # Max pixels to consider same ball
@@ -953,7 +959,7 @@ def _track_single_ball_candidate(
             kick_frame = frame
             break
     
-    return {
+    result = {
         "centers": centers,
         "boxes": boxes,
         "confs": confs,
@@ -969,6 +975,14 @@ def _track_single_ball_candidate(
         "has_kick": kick_frame is not None,
         "coverage": len(centers) / total if total else 0.0,
     }
+    
+    # Summary logging
+    kick_info = f"Kick @ frame {kick_frame}" if kick_frame else "No kick"
+    print(f"[_track_single_ball_candidate] Ball {candidate['id']} done: "
+          f"{len(centers)}/{total} frames ({result['coverage']:.0%}), "
+          f"max_vel={max_velocity:.1f}px/s, {kick_info}")
+    
+    return result
 
 
 def _detect_and_track_all_ball_candidates(