C3: Cross-View Cross-Modality Correspondence Dataset

Dataset for C3Po: Cross-View Cross-Modality Correspondence with Pointmap Prediction

arXiv | Project Website | GitHub

C3 contains 90K paired floor plans and photos from the Internet across 597 scenes with 153M pixel-level correspondences and 85K camera poses.

Image Pairs

image_pairs/ is split into train/, val, and test/, each with a image_pairs.csv.

• image_pairs.csv: Each row represents a plan-photo pair, consisting of uid, scene_name, plan_path, and photo_path. uid is used to reference corresponding files in correspondences/ and camera_poses, named using the format {int(uid):06d}.npy. scene_name is used to reference corresponding floor plans (visual/{scene_name}/{plan_path}) and photos (visual/{scene_name}/{photo_path}).

Correspondences and Camera Poses

geometric/ has three files: geometric_train.tar.gz, geometric_val.tar.gz, and geometric_test.tar.gz. Each of these files (geometric_{split}.tar.gz) can be extracted to {split}/correspondences/ and {split}/camera_poses/.

• correspondences/: Each .npy files contains an array of [plan_correspondences (M, 2), photo_correspondences (M, 2)] and are grouped in batches of 1,000.

• camera_poses/: Each .npy file contains an array of [R_plan-to-cam (3, 3), t_plan (3,), K (3, 3)] and are grouped in batches of 1,000.

geometric/
├── train/                                         # Extracted from geometric_train.tar.gz
│   ├── correspondences/                           
│   │   ├── 0/
│   │   │   ├── 00000.npy
│   │   │   ├── ...
│   │   │   ├── 00999.npy
│   │   ├── ...
│   ├── camera_poses/                              
│   │   ├── 0/
│   │   │   ├── 00000.npy
│   │   │   ├── ...
│   │   │   ├── 00999.npy
│   │   ├── ...
├── val/                                           # Extracted from geometric_val.tar.gz
│   ├── (same structure as train)
└── test/                                          # Extracted from geometric_test.tar.gz
    └── (same structure as train)

Floor Plans and Photos

visual/ contains floor plans and photos grouped by scenes.

visual/
├── Aachen_Cathedral.tar.gz
├── Abbatiale_d'Ottmarsheim.tar.gz
└── ...

Archived Contents

Each {scene_name}.tar.gz file contains the following structure when extracted:

├── images/
│   ├── commons/                                   # arbitrary number of subcategories
│   │   ├── {wikimedia_commons_subcategory_1}/     # arbitrary number of photos
│   │   │   ├── {photo_A}.png
│   │   │   ├── ...     
│   │   ├── {wikimedia_commons_subcategory_2}/
│   │   │   ├── ...
│   │   ├── ...
│   ├── flickr/                                    # arbitrary number of photos
│   │   ├── {photo_B}.png
│   │   ├── ...
├── plans/                                         # arbitrary number of floor plans
    ├── {floor_plan_A}.png
    ├── ...

Example Visualization

from os.path import join

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from PIL import Image

def draw_camera_frustum(ax, t_p, R_p2c, frustum_length, frustum_width, color='blue', alpha=0.3):
    # Camera axes
    forward = R_p2c.T[:, 2]  
    forward_xz = forward.copy()
    forward_xz[1] = 0                # Project onto the XZ plane

    if np.linalg.norm(forward_xz) < 1e-6:
        forward_xz = np.array([0, 0, 1])
    else:
        forward_xz /= np.linalg.norm(forward_xz)

    right_xz = np.cross(np.array([0, 1, 0]), forward_xz)
    right_xz /= np.linalg.norm(right_xz)

    # Near and far plane distances
    near_len, far_len = frustum_length * 0.2, frustum_length
    near_width, far_width = frustum_width * 0.2, frustum_width

    # Corner points of the frustum
    cc = -R_p2c.T @ t_p
    points = np.array([
        cc + forward_xz * near_len - right_xz * near_width / 2,     # near left
        cc + forward_xz * near_len + right_xz * near_width / 2,     # near right
        cc + forward_xz * far_len + right_xz * far_width / 2,       # far right
        cc + forward_xz * far_len - right_xz * far_width / 2.       # far left
    ])

    x, z = points[:, 0], points[:, 2]
    ax.fill(x, z, color=color, alpha=alpha)
    ax.plot(np.append(x, x[0]), np.append(z, z[0]), color=color)


# Load image pair
image_pairs_path = "image_pairs/train/image_pairs.csv"
image_pairs = pd.read_csv(image_pairs_path)
uid, scene_name, plan_path, photo_path = image_pairs.iloc[0]

# Load correspondences
geometric_train_dir = "geometric/train/"
corr_path = join(geometric_train_dir, "correspondences", f"{int(uid) // 1000}", f"{int(uid):06d}.npy")
plan_corr, photo_corr = np.load(corr_path)

# Load camera pose
camera_pose_path = join(geometric_train_dir, "camera_poses", f"{int(uid) // 1000}", f"{int(uid):06d}.npy")
R_p2c, t_p, _ = np.load(camera_pose_path, allow_pickle=True)
R_p2c = np.array(R_p2c.tolist(), dtype=float)
t_p = np.array(t_p)

# Load floor plan and photo
visual_dir = "visual/"
plan = Image.open(join(visual_dir, scene_name, plan_path)).convert("RGB")
photo = Image.open(join(visual_dir, scene_name, photo_path)).convert("RGB")

# Visualize
fig, axes = plt.subplots(1, 2, figsize=(12, 6))
fig.suptitle(f"Scene name: {scene_name}", fontsize=16)

axes[0].imshow(plan)
axes[0].set_title("Floor Plan")
axes[0].scatter(plan_corr[:, 0], plan_corr[:, 1], c="r", s=1)
scale = max(plan.size) * 0.05
draw_camera_frustum(axes[0], t_p, R_p2c, frustum_length=scale, frustum_width=scale, color='blue', alpha=0.3)
axes[0].axis('off')

axes[1].imshow(photo)
axes[1].set_title("Photo")
axes[1].scatter(photo_corr[:, 0], photo_corr[:, 1], c="r", s=1)
axes[1].axis('off')

plt.tight_layout()
plt.show()

Citation

If you use data from C3, please cite with the following:

@inproceedings{
      huang2025c3po,
      title={C3Po: Cross-View Cross-Modality Correspondence by Pointmap Prediction}, 
      author={Huang, Kuan Wei and Li, Brandon and Hariharan, Bharath and Snavely, Noah},
      booktitle={Advances in Neural Information Processing Systems},
      volume={38},
      year={2025}
}