tools/analyze_spatial_relationships.py

from smolagents import tool
import os
import json
import math
import cv2
import numpy as np
from typing import Optional

@tool
def analyze_spatial_relationships(image_path: Optional[str] = None, object_type1: Optional[str] = None, object_type2: Optional[str] = None) -> str:
    """Analyze spatial relationships between objects in an image.
    Args:
        image_path: Path to the image file to analyze.
        object_type1: First object type to analyze.
        object_type2: Second object type to analyze.
    Returns:
        A description of spatial relationships between the objects.
    """
    try:
        # Check if image path exists
        if not image_path or not os.path.exists(image_path):
            return f"Error: Image path '{image_path}' does not exist."

        # Check if detection data exists, if not run detection
        detection_data_path = f"{os.path.splitext(image_path)[0]}_detection_data.json"

        if not os.path.exists(detection_data_path):
            # Run detection first
            from tools.detect_objects import detect_objects_with_roboflow
            detect_objects_with_roboflow(image_path=image_path)

            if not os.path.exists(detection_data_path):
                return f"Error: Detection failed. Could not create detection data."

        # Load existing detection data
        with open(detection_data_path, "r") as f:
            result = json.load(f)

        # Load the original image to get dimensions
        img = cv2.imread(image_path)
        if img is None:
            return f"Error: Could not read image at '{image_path}'."

        img_height, img_width = img.shape[:2]

        # Define "close" threshold as 30% of image width
        close_threshold = img_width * 0.3

        # Extract objects of the specified types
        objects1 = []
        objects2 = []

        if "predictions" in result:
            for prediction in result["predictions"]:
                class_name = prediction["class"]

                # Check if this prediction matches either object type
                if object_type1 and object_type1.lower() in class_name.lower():
                    if "x" in prediction and "y" in prediction:
                        objects1.append({
                            "class": class_name,
                            "x": prediction["x"],
                            "y": prediction["y"],
                            "width": prediction["width"],
                            "height": prediction["height"],
                            "confidence": prediction["confidence"]
                        })

                if object_type2 and object_type2.lower() in class_name.lower():
                    if "x" in prediction and "y" in prediction:
                        objects2.append({
                            "class": class_name,
                            "x": prediction["x"],
                            "y": prediction["y"],
                            "width": prediction["width"],
                            "height": prediction["height"],
                            "confidence": prediction["confidence"]
                        })

        # If no objects found, return appropriate message
        if not objects1:
            return f"No objects matching '{object_type1}' were found in the image."

        if not objects2:
            return f"No objects matching '{object_type2}' were found in the image."

        # Analyze spatial relationships
        close_pairs = []

        for obj1 in objects1:
            for obj2 in objects2:
                # Calculate Euclidean distance between centers
                distance = math.sqrt(
                    (obj1["x"] - obj2["x"])**2 +
                    (obj1["y"] - obj2["y"])**2
                )

                # Check if they are close
                if distance <= close_threshold:
                    close_pairs.append({
                        "object1": obj1["class"],
                        "object2": obj2["class"],
                        "distance": distance,
                        "distance_percent": (distance / img_width) * 100
                    })

        # Format the results
        if close_pairs:
            output = f"Found {len(close_pairs)} instances where {object_type1} is close to {object_type2}:\n"
            for i, pair in enumerate(close_pairs, 1):
                output += f"{i}. {pair['object1']} is close to {pair['object2']} (distance: {pair['distance']:.1f} pixels, {pair['distance_percent']:.1f}% of image width)\n"

            output += f"\nClose is defined as objects with centers within {close_threshold:.1f} pixels (30% of image width) from each other."
            return output
        else:
            return f"No instances found where {object_type1} is close to {object_type2}. Close is defined as objects with centers within {close_threshold:.1f} pixels (30% of image width) from each other."

    except Exception as e:
        return f"Error analyzing spatial relationships: {str(e)}"