feat: add data preparation script for YOLOv8

prepare_cheerios_yolov8.py

@@ -0,0 +1,126 @@
+import os
+import numpy as np
+import shutil
+from pathlib import Path
+from scipy.spatial.transform import Rotation as R
+import cv2
+
+# --- Configuration ---
+SOURCE_DATA_DIR = Path('/home/pik/dev/hackwhyd')
+YOLO_DIR = SOURCE_DATA_DIR / 'YOLOv8'
+DEST_DATA_DIR = YOLO_DIR / 'data' / 'cheerios'
+CLASS_NAME = 'cheerios'
+CLASS_ID = 0
+
+# --- Load Camera Intrinsics ---
+k_matrix = np.load(SOURCE_DATA_DIR / 'k_matrix.npy')
+
+# --- Image Dimensions (assuming all images are the same size) ---
+IMG_WIDTH = 1920
+IMG_HEIGHT = 1080
+
+# --- Create Destination Directories ---
+DEST_IMAGES_DIR = DEST_DATA_DIR / 'images'
+DEST_LABELS_DIR = DEST_DATA_DIR / 'labels'
+DEST_IMAGES_DIR.mkdir(parents=True, exist_ok=True)
+DEST_LABELS_DIR.mkdir(parents=True, exist_ok=True)
+
+def project_to_2d(p, q, k_matrix):
+    # (x,y,z) -> (y,-z,x) is the transformation from camera to view (the frame of reference you do projections in)
+    l = 15
+    r = R.from_quat(q)
+    x = p + r.apply(np.array([l, 0, 0]))
+    y = p + r.apply(np.array([0, l, 0]))
+    z = p + r.apply(np.array([0, 0, l]))
+    x = np.array([x[1], -x[2], x[0]])
+    y = np.array([y[1], -y[2], y[0]])
+    z = np.array([z[1], -z[2], z[0]])
+    x = np.dot(k_matrix, x)
+    y = np.dot(k_matrix, y)
+    z = np.dot(k_matrix, z)
+    x /= x[2]
+    y /= y[2]
+    z /= z[2]
+    return x, y, z
+
+def process_dataset_split(split_name):
+    """Processes a dataset split (e.g., 'train', 'val')."""
+    source_split_dir = SOURCE_DATA_DIR / split_name
+    source_labels_dir = source_split_dir / 'labels'
+    source_images_dir = source_split_dir / 'images'
+
+    if not source_labels_dir.exists():
+        print(f"Warning: Label directory not found for split '{split_name}': {source_labels_dir}")
+        return
+
+    for label_file in sorted(source_labels_dir.glob('*.txt')):
+        # --- Read Original Label ---
+        with open(label_file, 'r') as f:
+            original_line = f.readline().strip()
+        
+        parts = original_line.split()
+        
+        p = np.asarray([float(parts[3]), float(parts[4]), float(parts[5])])
+        q = np.asarray([float(parts[6]), float(parts[7]), float(parts[8]), float(parts[9])])
+
+        x, y, z = project_to_2d(p, q, k_matrix)
+
+        x_coords = [x[0], y[0], z[0]]
+        y_coords = [x[1], y[1], z[1]]
+
+        x_min = min(x_coords)
+        y_min = min(y_coords)
+        x_max = max(x_coords)
+        y_max = max(y_coords)
+
+        x_center = (x_min + x_max) / 2
+        y_center = (y_min + y_max) / 2
+        width = x_max - x_min
+        height = y_max - y_min
+
+        # Normalize
+        x_center /= IMG_WIDTH
+        y_center /= IMG_HEIGHT
+        width /= IMG_WIDTH
+        height /= IMG_HEIGHT
+
+        # --- Write New Label File ---
+        dest_label_file = DEST_LABELS_DIR / label_file.name
+        with open(dest_label_file, 'w') as f:
+            f.write(f"{CLASS_ID} {x_center} {y_center} {width} {height}\n")
+
+        # --- Copy Image ---
+        image_name = label_file.stem + '.jpg'
+        source_image_path = source_images_dir / image_name
+        dest_image_path = DEST_IMAGES_DIR / image_name
+        if source_image_path.exists():
+            shutil.copy(source_image_path, dest_image_path)
+        else:
+            print(f"Warning: Image not found: {source_image_path}")
+
+# --- Process Train and Validation Sets ---
+print("Processing train set...")
+process_dataset_split('train')
+print("Processing val set...")
+process_dataset_split('val')
+
+# --- Create cheerios.yaml ---
+yaml_path = YOLO_DIR / 'cheerios_yolov8.yaml'
+
+yaml_content = f"""
+path: {DEST_DATA_DIR.resolve()}
+train: images
+val: images
+
+# number of classes
+nc: 1
+
+# class names
+names: ['{CLASS_NAME}']
+"""
+with open(yaml_path, 'w') as f:
+    f.write(yaml_content)
+
+print("\nDataset preparation complete.")
+print(f"New dataset created at: {DEST_DATA_DIR}")
+print(f"Config file created at: {yaml_path}")