Implemented coarse and granular matching

src/label_episodes.py

@@ -141,15 +141,55 @@ def match_episodes_to_references(episodes: list[str], references: dict[str, list
 
 
 def _match_episode_batch(
-    frames: list[np.ndarray], references: dict[str, list[np.ndarray]], step_size: int
+    frames: list[np.ndarray],
+    references: dict[str, list[np.ndarray]],
+    step_size: int,
+    significance_threshold: float = 0.8,
 ) -> dict[str, float]:
-    """Worker function: compare a batch of frames to all references, return local max per reference."""
+    """
+    Worker function: compare a batch of frames to all references, return local max per reference.
+
+    Strategy:
+    - Compare every `step_size` frame.
+    - If a score >= significance_threshold is found, check the +/- (step_size - 1) neighborhood
+      to refine the max score.
+    """
     local_results = {ref: 0.0 for ref in references.keys()}
-    for frame in frames:
-        for ref, ref_images in references.items():
-            for ref_img in ref_images:
-                score = compare_images(frame, ref_img)
-                local_results[ref] = max(local_results[ref], score)
+    n_frames = len(frames)
+    significant_coarse_match: int | None = None
+
+    for ref, ref_images in references.items():
+        for ref_img in ref_images:
+            frame_index = 0
+            while frame_index < n_frames:
+                if significant_coarse_match == frame_index:
+                    frame_index += 1  # Skip previously matched frame
+                    continue
+
+                # Score current frame
+                score = compare_images(frames[frame_index], ref_img)
+                if score > local_results[ref]:
+                    local_results[ref] = score
+
+                # Handle granular matching
+                if significant_coarse_match:
+                    if frame_index >= significant_coarse_match + step_size:
+                        if score < significance_threshold:
+                            # End of granular matching
+                            significant_coarse_match = None
+                            frame_index += step_size
+                            continue
+                        else:
+                            # Continue granular matching, but don't recheck previous section
+                            significant_coarse_match = frame_index
+
+                    frame_index += 1  # Continue granular matching
+                elif score >= significance_threshold:
+                    frame_index -= step_size + 1  # Check previous frames
+                    significant_coarse_match = frame_index  # Mark significant match
+                else:
+                    frame_index += step_size
+
     return local_results