first refactored version

createDataset.py

@@ -6,7 +6,7 @@ from urllib.parse import urlparse
 from tqdm.auto import tqdm
 
 
-# --- KONFIGURATION ---
+# s3 bucket configuration
 MINIO_CONFIG = {
     'endpoint_url': 'https://minio.hgk.ch',
     'access_key': 'meinAccessKey',
@@ -14,28 +14,32 @@ MINIO_CONFIG = {
     'bucket': 'skiai'
 }
 
+# input specs, annotations
 JSON_PATH = 'datasets/skier_pose/labelstudio_export.json'
-OUTPUT_DIR = 'datasets/skier_pose'
+# input specs, keypoint orde must stay consistent
-TRAIN_RATIO = 0.8
-
-# Die Reihenfolge MUSS konsistent bleiben
 KP_ORDER = [
     "leftski_tip", "leftski_tail", "rightski_tip", "rightski_tail",
     "leftpole_top", "leftpole_bottom", "rightpole_top", "rightpole_bottom"
 ]
 
+# output specs
+OUTPUT_DIR = 'datasets/skier_pose'
+TRAIN_RATIO = 0.8
+
+
+# create folder structure
+def __setup_directories():
 
-def setup_directories():
-    """Erstellt die Struktur: train/images, train/labels, val/images, val/labels"""
     for split in ['train', 'val']:
         os.makedirs(os.path.join(OUTPUT_DIR, split, 'images'), exist_ok=True)
         os.makedirs(os.path.join(OUTPUT_DIR, split, 'labels'), exist_ok=True)
 
 
-def download_from_minio(s3_path, local_path):
+# download image from s3
+def __download_from_minio(s3_path, local_path):
     parsed = urlparse(s3_path)
     bucket = MINIO_CONFIG['bucket']
-    # Entfernt 's3://bucketname/' falls vorhanden, sonst nur den slash
+    # removes 's3://bucketname/' if existing, otherwise slash
     key = parsed.path.lstrip('/')
 
     s3 = boto3.client('s3',
@@ -45,8 +49,11 @@ def download_from_minio(s3_path, local_path):
     s3.download_file(bucket, key, local_path)
 
 
-def convert_to_yolo():
+# create YOLO dataset
-    setup_directories()
+def createYOLOdataset():
+    __setup_directories()
+
+    # read annotations
     with open(JSON_PATH, 'r', encoding='utf-8') as f:
         data = json.load(f)
 
@@ -54,10 +61,11 @@ def convert_to_yolo():
     random.shuffle(data)
     split_idx = int(len(data) * TRAIN_RATIO)
 
+    # loop over all images
     for i, entry in enumerate(tqdm(data, desc="Importing Images", unit="img")):
         split = 'train' if i < split_idx else 'val'
 
-        # Dateinamen aus dem 'data'-Feld holen
+        # get image name
         image_s3_path = entry['data']['image']
         filename = os.path.basename(image_s3_path)
         base_name = os.path.splitext(filename)[0]
@@ -66,20 +74,20 @@ def convert_to_yolo():
         label_local_path = os.path.join(OUTPUT_DIR, split, 'labels', f"{base_name}.txt")
 
         try:
-            download_from_minio(image_s3_path, img_local_path)
+            __download_from_minio(image_s3_path, img_local_path)
         except Exception as e:
             tqdm.write(f"Error treating {filename}: {e}")
             continue
 
         yolo_lines = []
 
-        # Sicherstellen, dass Annotationen vorhanden sind
+        # check if annotations, otherwise skip
         if not entry.get('annotations'):
             continue
 
         results = entry['annotations'][0].get('result', [])
 
-        # Hilfsmaps für das Matching über IDs
+        # dummy vars
         kp_map = {}  # ID -> {label, x, y}
         visibility_map = {}  # ID -> v_status (1 oder 2)
         bboxes = []  # Liste aller gefundenen BBoxes
@@ -97,8 +105,6 @@ def convert_to_yolo():
                 }
 
             elif res_type == 'choices':
-                # Matching: Label Studio nutzt die gleiche ID für Keypoint und Choice
-                # Wir prüfen, ob die Checkbox "1" (dein Alias für verdeckt) gewählt wurde
                 if "1" in val.get('choices', []):
                     visibility_map[res_id] = 1
 
@@ -110,18 +116,16 @@ def convert_to_yolo():
                 by = (val['y'] / 100.0) + (bh / 2.0)
                 bboxes.append(f"{bx:.6f} {by:.6f} {bw:.6f} {bh:.6f}")
 
-        # Für jede gefundene BBox eine YOLO Zeile generieren
+        # create yolo data
-        # (Hinweis: Aktuell werden alle Keypoints an jede BBox gehängt)
         for bbox_coords in bboxes:
             line = f"0 {bbox_coords}"
 
             for kp_name in KP_ORDER:
-                # Finde die ID des Keypoints mit diesem Namen
                 target_id = next((id for id, d in kp_map.items() if d['label'] == kp_name), None)
 
                 if target_id:
                     coords = kp_map[target_id]
-                    # Sichtbarkeit: 1 (verdeckt) wenn in visibility_map, sonst 2 (sichtbar)
+                    # visibility, 0 missing, 1 invisible, 2 visible
                     v = visibility_map.get(target_id, 2)
                     line += f" {coords['x']:.6f} {coords['y']:.6f} {v}"
                 else:
@@ -132,7 +136,7 @@ def convert_to_yolo():
         with open(label_local_path, 'w', encoding='utf-8') as f:
             f.write('\n'.join(yolo_lines))
 
-    print(f"Fertig! Daten liegen in: {os.path.abspath(OUTPUT_DIR)}")
+    print(f"Finished! Dataset saved to: {os.path.abspath(OUTPUT_DIR)}")
 
 if __name__ == "__main__":
-    convert_to_yolo()
+    createYOLOdataset()

main.py

@@ -1,16 +1,18 @@
+
+
 from ultralytics import YOLO
 
 # Create a new YOLO model from scratch
 #model = YOLO("yolo11n.yaml")
 
 # Load a pretrained YOLO model (recommended for training)
-model = YOLO("yolo11n-pose.pt")
+#model = YOLO("yolo11n-pose.pt")
 
 # Train the model using the 'coco8.yaml' dataset for 3 epochs
-results = model.train(data="datasets/skier_pose/skier_pose.yaml", epochs=200, imgsz=640, device='mps')
+#results = model.train(data="datasets/skier_pose/skier_pose.yaml", epochs=200, imgsz=640, device='mps')
 
 # Evaluate the model's performance on the validation set
-results = model.val()
+#results = model.val()
 
 # Perform object detection on an image using the model
 #results = model("bus.jpg", device='cpu')

specs/skier_pose.yaml

@@ -0,0 +1,24 @@
+path: datasets/skier_pose  # relativer Pfad zum Dataset-Ordner
+train: train
+val: val
+
+# Keypoints Konfiguration
+kpt_shape: [8, 3]  # 8 Keypoints, jeweils (x, y, visibility)
+
+# Flip-Indizes für Augmentation (Spiegelung)
+# Wenn das Bild gespiegelt wird:
+# 0 (left tip) wird zu 2 (right tip) und umgekehrt
+# 1 (left tail) wird zu 3 (right tail) und umgekehrt
+# 4 (left pole top) wird zu 6 (right pole top) etc.
+flip_idx: [2, 3, 0, 1, 6, 7, 4, 5]
+
+# Skeleton: Verbindungen für die Visualisierung und Strukturhilfe
+# Definiert Paare von Keypoint-Indizes (0-basiert)
+skeleton:
+  - [0, 1] # Linker Ski (Tip zu Tail)
+  - [2, 3] # Rechter Ski (Tip zu Tail)
+  - [4, 5] # Linker Stock (Top zu Bottom)
+  - [6, 7] # Rechter Stock (Top zu Bottom)
+
+names:
+  0: skier