script.py

import os
import pandas as pd
from ultralytics import YOLO
import glob

# --- CONFIGURATION ---
TEST_IMAGE_PATH = "/tmp/data/test_images"
SUBMISSION_SAVE_PATH = "submission.csv"
MODEL_WEIGHTS = "best.pt"

# ✅ 1. Balanced Confidence
# 0.10 captures weaker tools but keeps precision reasonable.
CONF_THRESHOLD = 0.10

def get_category_id(cls_id):
    # IDs 0, 1, 2 are correct.
    return int(cls_id)

def run_inference(model, image_path, conf_threshold, save_path):
    print(f"🚀 Checking for images in {image_path}...")
    
    test_images = []
    if os.path.exists(image_path):
        extensions = ['*.png', '*.jpg', '*.jpeg', '*.bmp']
        for ext in extensions:
            test_images.extend(glob.glob(os.path.join(image_path, ext)))
        test_images.sort()
        test_images = [os.path.basename(x) for x in test_images]
    else:
        print(f"⚠️ Warning: {image_path} not found.")

    print(f"🔍 Found {len(test_images)} images.")
    df_rows = []

    if len(test_images) > 0:
        for image_name in test_images:
            full_path = os.path.join(image_path, image_name)
            
            # 🚀 BALANCED TTA SETTINGS
            results = model.predict(
                full_path, 
                conf=conf_threshold, 
                imgsz=1024,       # Keep High Res
                augment=True,     # ⬅️ ENABLED: Test Time Augmentation (Boosts accuracy)
                iou=0.60,         # Merge overlapping boxes
                max_det=30,       # ⬅️ INCREASED: 5 was too strict, 30 is safe
                verbose=False
            )
            
            bbox_list = []
            category_list = []
            
            for result in results:
                for box in result.boxes:
                    # YOLO (Center) -> COCO (Top-Left)
                    x_c, y_c, w, h = box.xywh[0].tolist()
                    x_min = x_c - (w / 2)
                    y_min = y_c - (h / 2)
                    
                    bbox_list.append([x_min, y_min, w, h])
                    
                    cls_id = int(box.cls[0])
                    category_list.append(get_category_id(cls_id))

            df_rows.append({
                "file_name": image_name,
                "bbox": str(bbox_list),
                "category_id": str(category_list)
            })

    df_predictions = pd.DataFrame(df_rows, columns=["file_name", "bbox", "category_id"])
    if df_predictions.empty:
        df_predictions = pd.DataFrame(columns=["file_name", "bbox", "category_id"])

    df_predictions.to_csv(save_path, index=False)
    print("✅ Done!")

if __name__ == "__main__":
    try:
        print(f"🔥 Loading Model...")
        model = YOLO(MODEL_WEIGHTS)
        run_inference(model, TEST_IMAGE_PATH, CONF_THRESHOLD, SUBMISSION_SAVE_PATH)
    except Exception as e:
        print(f"❌ Error: {e}")