Upload api.py

api.py

@@ -1012,7 +1012,8 @@ async def detect_scenes(
 
     X = np.array(frames)
     labels = hierarchical_cluster_with_min_size(X, max_groups, min_cluster_size).tolist()
-    print(f"Scene clustering jeràrquic: {len(set([l for l in labels if l >= 0]))} clusters")
+    initial_clusters = len(set([l for l in labels if l >= 0]))
+    print(f"Scene clustering jeràrquic inicial: {initial_clusters} clusters")
 
     # Agrupar per etiqueta (>=0)
     clusters = {}
@@ -1020,6 +1021,59 @@ async def detect_scenes(
         if lbl is None or lbl < 0:
             continue
         clusters.setdefault(int(lbl), []).append(i)
+    
+    # VALIDACIÓ: Mesurar robustesa dels clusters i fusionar si són massa similars
+    # Calcular centroides (histograma promig de cada cluster)
+    centroids = {}
+    for lbl, idxs in clusters.items():
+        cluster_histograms = X[idxs]
+        centroids[lbl] = np.mean(cluster_histograms, axis=0)
+    
+    # Comparar distàncies entre clusters
+    # Si dos clusters tenen una distància euclidiana < threshold, són massa similars
+    SIMILARITY_THRESHOLD = 0.15  # Ajustable: més baix = més estricte
+    
+    # Calcular matriu de distàncies entre centroides
+    cluster_labels = sorted(centroids.keys())
+    distances = {}
+    for i, lbl1 in enumerate(cluster_labels):
+        for lbl2 in cluster_labels[i+1:]:
+            dist = np.linalg.norm(centroids[lbl1] - centroids[lbl2])
+            distances[(lbl1, lbl2)] = dist
+    
+    # Trobar parelles de clusters massa similars i fusionar-los
+    merged = {}  # mapatge de label_old -> label_new
+    for lbl in cluster_labels:
+        merged[lbl] = lbl
+    
+    # Fusionar clusters similars (greedy approach)
+    for (lbl1, lbl2), dist in sorted(distances.items(), key=lambda x: x[1]):
+        if dist < SIMILARITY_THRESHOLD:
+            # Fusionar lbl2 amb lbl1
+            current_lbl1 = merged.get(lbl1, lbl1)
+            current_lbl2 = merged.get(lbl2, lbl2)
+            if current_lbl1 != current_lbl2:
+                # Assignar lbl2 al grup de lbl1
+                for k, v in merged.items():
+                    if v == current_lbl2:
+                        merged[k] = current_lbl1
+                print(f"[SCENE VALIDATION] Fusionant clusters {lbl2} i {lbl1} (distància={dist:.3f})")
+    
+    # Aplicar fusió als clusters
+    new_clusters = {}
+    for lbl, idxs in clusters.items():
+        new_lbl = merged[lbl]
+        if new_lbl not in new_clusters:
+            new_clusters[new_lbl] = []
+        new_clusters[new_lbl].extend(idxs)
+    
+    clusters = new_clusters
+    final_clusters = len(clusters)
+    eliminated = initial_clusters - final_clusters
+    
+    if eliminated > 0:
+        print(f"[SCENE VALIDATION] Reduït de {initial_clusters} a {final_clusters} clusters "
+              f"(eliminats {eliminated} clusters massa similars)")
 
     # Escriure imatges representatives per a cada clúster
     base = TEMP_ROOT / video_name / "scenes"