| # Real-World Evaluation Images for Articulated Objects Interaction Generation | |
| This dataset contains the real-world images used in evaluating [DragAPart](https://dragapart.github.io/), a conditional image generator that models interaction with articulated objects. | |
| ## 📦 How to Use It? | |
| Each sample consists of: | |
| - `original_image_XXX.png`: The base image showing an articulated object. | |
| - `arrow_locations_XXX.npy`: A NumPy file containing the arrow coordinates for interaction. | |
| The `.npy` file stores one arrow as: | |
| ```python | |
| [x0, y0, x1, y1] # Normalized coordinates in [0, 1] | |
| ``` | |
| Where: | |
| - `(x0, y0)` is the **starting point** of the interaction (e.g., where the user clicks), | |
| - `(x1, y1)` is the **end point** indicating the direction or extent of the manipulation. | |
| These coordinates are normalized relative to the image size. | |
| --- | |
| ## 🖼️ Visualization | |
| You can visualize the interaction using the following Python script: | |
| ```python | |
| import numpy as np | |
| from PIL import Image | |
| import matplotlib.pyplot as plt | |
| # Load image and arrow data | |
| image_path = "original_image_000.png" | |
| arrow_path = "arrow_locations_000.npy" | |
| image = Image.open(image_path) | |
| arrow = np.load(arrow_path)[0] # [x0, y0, x1, y1] | |
| # Convert normalized coordinates to pixel values | |
| width, height = image.size | |
| x0, y0 = int(arrow[0] * width), int(arrow[1] * height) | |
| x1, y1 = int(arrow[2] * width), int(arrow[3] * height) | |
| # Plot the image and overlay the interaction arrow | |
| plt.figure(figsize=(6, 6)) | |
| plt.imshow(image) | |
| plt.arrow(x0, y0, x1 - x0, y1 - y0, | |
| color='red', width=2, head_width=10, length_includes_head=True) | |
| plt.axis('off') | |
| plt.title("Interactive Manipulation Arrow") | |
| plt.show() | |
| ``` | |
| This will display the original image with a red arrow showing the suggested user interaction as below: | |
|  | |
|  | |