Question about Ground Truth Alignment Error in Venman Dataset

[minwoo0611]

Dear Authors,

First, thank you for sharing the valuable Wild-Places dataset. It has been incredibly helpful for research in forestry, particularly for localization and place recognition tasks.

I have a question regarding the ground truth data in the Venman dataset, specifically for V-02 and V-03 (I haven't tested other sequences yet).

I created a global map using the code below:

import open3d as o3d
import pandas as df
import numpy as np
import os


def read_poses(csv_path):
    """Read poses from CSV file with string timestamps."""
    poses_df = df.read_csv(csv_path, dtype={'timestamp': str})

    poses = {}
    for _, row in poses_df.iterrows():
        timestamp = row['timestamp']  
        print(timestamp)  
        translation = np.array([row['x'], row['y'], row['z']])
        quaternion = np.array([row['qw'], row['qx'], row['qy'], row['qz']]) 
        poses[timestamp] = {'translation': translation, 'quaternion': quaternion}
    return poses


def quaternion_to_rotation_matrix(q):
    """Convert quaternion (w, x, y, z) to 4x4 transformation matrix."""
    w, x, y, z = q
    r = np.array([
        [1 - 2*y*y - 2*z*z, 2*x*y - 2*z*w, 2*x*z + 2*y*w, 0],
        [2*x*y + 2*z*w, 1 - 2*x*x - 2*z*z, 2*y*z - 2*x*w, 0],
        [2*x*z - 2*y*w, 2*y*z + 2*x*w, 1 - 2*x*x - 2*y*y, 0],
        [0, 0, 0, 1]
    ])
    return r

def combine_point_clouds(pcd_dir, csv_path):
    """Combine point clouds into a global map using poses."""
    # Read poses
    poses = read_poses(csv_path)
    
    # Initialize global point cloud
    global_pcd = o3d.geometry.PointCloud()
    i = 0
    # Process each point cloud
    for filename in os.listdir(pcd_dir):
        
        if filename.endswith('.pcd'):
            # Extract timestamp from filename
            timestamp = str(filename[:-4])  # Remove '.pcd' extension
            if timestamp in poses:
                # Read point cloud
                pcd_path = os.path.join(pcd_dir, filename)
                pcd = o3d.io.read_point_cloud(pcd_path)
                
                # Get pose
                pose = poses[timestamp]
                translation = pose['translation']
                quaternion = pose['quaternion']
                
                # Create transformation matrix
                transform = np.eye(4)
                transform[:3, :3] = quaternion_to_rotation_matrix(quaternion)[:3, :3]
                transform[:3, 3] = translation
                
                # Transform point cloud
                pcd.transform(transform)
                
                # Add to global point cloud
                global_pcd += pcd
                i = i+1
            else:
                print(f"Warning: No pose found for timestamp {timestamp}")
    print(i)
    return global_pcd

def main():
    # Define paths
    pcd_dir = "V-03/Clouds_downsampled"
    csv_path = "V-03/poses_aligned.csv"
    
    # Combine point clouds
    global_map = combine_point_clouds(pcd_dir, csv_path)
    
    # Save global map
    output_path = "global_map_v03.pcd"
    o3d.io.write_point_cloud(output_path, global_map)
    print(f"Global map saved to {output_path}")
    

if __name__ == "__main__":
    main()

When visualizing the resulting global map, I noticed significant misalignment between the point clouds from V-02 (green) and V-03 (red), particularly for features like streetlights. The output suggests that these features are not properly aligned.

My question is, Should I expect this level of ground truth error in the dataset, or is there potentially an issue in my global map generation process? I used the provided data (point clouds and poses) as-is, as shown in the code above.

I would appreciate your insights on whether this misalignment is due to inherent ground truth errors or a mistake in my implementation.

Thank you for your time and support!

Best Regards,
Minwoo