Upload folder using huggingface_hub

app.py

@@ -28,9 +28,21 @@ DROP_WORDS = {
     "image", "photo", "picture", "view", "scene", "background", "foreground",
     "left", "right", "top", "bottom", "front", "back", "side", "area", "part",
     "visible", "large", "small", "several", "multiple", "many", "some",
+    "is", "are", "was", "were", "be", "being", "been", "appears", "appear",
+    "of", "in", "it", "its", "to", "for", "by", "as", "at", "from", "into",
+    "parked", "parking", "surrounded", "including", "taken", "shining",
+    "brightly", "different", "models", "colors", "color", "angle", "high",
+    "panoramic", "modern", "curved", "few", "blue", "red", "white",
+}
+KEEP_PHRASES = {
+    "parking lot",
+}
+DROP_LABELS = {
+    "roof", "floor", "floors", "balcony", "balconies", "sun", "cloud", "clouds",
+    "sky", "brightly", "angle", "high angle",
 }
 CAPTION_SPLIT = re.compile(
-    r"[,.;:]|\\bwith\\b|\\band\\b|\\bnext to\\b|\\bin front of\\b|\\bbehind\\b|\\bon\\b|\\balong\\b|\\bnear\\b",
+    r"[,.;:]|\\bwith\\b|\\band\\b|\\bnext to\\b|\\bin front of\\b|\\bbehind\\b|\\bon\\b|\\balong\\b|\\bnear\\b|\\bsurrounded by\\b",
     flags=re.IGNORECASE,
 )
 
@@ -39,12 +51,23 @@ def _clean_label(value: str) -> str:
     text = str(value).strip().lower()
     text = re.sub(r"[_/\\-]+", " ", text)
     text = re.sub(r"[^a-z0-9\\s]+", " ", text)
+    for phrase in KEEP_PHRASES:
+        if phrase in text:
+            return phrase
     words = [w for w in text.split() if w and w not in DROP_WORDS]
     if not words:
         return ""
+    if len(words) == 1:
+        word = words[0]
+        if len(word) > 3 and word.endswith("ies"):
+            word = word[:-3] + "y"
+        elif len(word) > 3 and word.endswith("s") and not word.endswith("ss"):
+            word = word[:-1]
+        words = [word]
     if len(words) > 5:
         words = words[-5:]
-    return " ".join(words)
+    label = " ".join(words)
+    return "" if label in DROP_LABELS else label
 
 
 def _labels_from_caption(text: str) -> list[str]:
@@ -54,12 +77,13 @@ def _labels_from_caption(text: str) -> list[str]:
         if not clean:
             continue
         words = clean.split()
-        # Prefer compact noun-like endings from descriptive chunks while keeping
-        # short labels intact. This is intentionally generic, not a fixed class
-        # list for one dataset.
-        for candidate in (" ".join(words[-3:]), " ".join(words[-2:]), words[-1]):
+        candidates = [clean]
+        if len(words) >= 3:
+            candidates.append(" ".join(words[-2:]))
+        candidates.append(words[-1])
+        for candidate in candidates:
             candidate = _clean_label(candidate)
-            if candidate and len(candidate) > 2:
+            if candidate and len(candidate) > 2 and candidate not in labels:
                 labels.append(candidate)
                 break
     return labels