README.md

Dataset for Aerial Vision-and-Dialog Navigation (AVDN)

This repository contains the Aerial Vision-and-Dialog Navigation (AVDN) dataset, proposed in the paper "Aerial Vision-and-Dialog Navigation." The dataset and associated tasks enable training and evaluation of models designed for navigation tasks guided by visual and dialog-based cues.

• Paper: Aerial Vision-and-Dialog Navigation
• Project Webpage: Aerial Vision-and-Dialog Navigation Project

Tasks

We introduce two tasks within the AVDN framework:

ANDH task

The ANDH task involves sub-trajectory data, where navigation occurs over smaller segments of the overall route. Instructions for downloading the dataset related to this task are available on the project webpage.

ANDH-Full task

The ANDH-Full task uses the full trajectory data, providing a more comprehensive view of the navigation route, from start to destination, which is shown in this repo .

Dataset format

The data in the CSV files provided in this repository represents the full trajectory data before it is split by dialog turns. The CSV format includes detailed information about each trajectory, including navigation instructions, GPS data, and path-related information.

To use this data for training and inference according to our current vision, you'll need to convert it from CSV format to JSON format. Below is a script that converts the CSV data into the required JSON format.

CSV to JSON conversion script

import pandas as pd
import json

# Process each CSV file split
for split in ['train', 'val_seen', 'val_unseen', 'test_unseen']:
    # Load the CSV data
    df = pd.read_csv(f'./{split}_full_data.csv')
    
    # Initialize a list to hold JSON data
    json_data = []
    
    # Convert each row in the DataFrame back to a dictionary (similar to JSON structure)
    for _, row in df.iterrows():
        entry = {
            'pre_dialogs': [],
            'map_name': row['map_name'],
            'route_index': row['route_index'],
            'instructions': row['instructions'],
            'gps_botm_left': json.loads(row['gps_botm_left']) if isinstance(row['gps_botm_left'], str) else row['gps_botm_left'],
            'gps_top_right': json.loads(row['gps_top_right']) if isinstance(row['gps_top_right'], str) else row['gps_top_right'],
            'lng_ratio': float(row['lng_ratio']),
            'lat_ratio': float(row['lat_ratio']),
            'angle': row['angle'],
            'destination': json.loads(row['destination']) if isinstance(row['destination'], str) else row['destination'],
            'attention_list': json.loads(row['attention_list']) if isinstance(row['attention_list'], str) else row['attention_list'],
            'gt_path_corners': json.loads(row['gt_path_corners']) if isinstance(row['gt_path_corners'], str) else row['gt_path_corners']
        }
        json_data.append(entry)

    # Save the data to a JSON file
    json_output_path = f'./{split}_full_data_from_csv.json'
    with open(json_output_path, 'w') as f:
        json.dump(json_data, f, indent=4)

Explanations of the key fields in the dataset

Below are explanations of the key fields in the dataset, with all coordinates provided in the format [Latitude, Longitude]:

• map_name: The name of the satellite image file corresponding to the environment in which the navigation occurs.
• route_index: The index of the specific trajectory being followed by the agent within the map.
• instructions: Step-by-step natural language instructions guiding the agent along the route.
• gps_botm_left: GPS coordinates representing the bottom-left corner of the map's bounding box.
• gps_top_right: GPS coordinates representing the top-right corner of the map's bounding box.
• lng_ratio: The ratio used to scale the map's longitude (horizontal) distance to the corresponding pixel dimensions in the image.
• lat_ratio: The ratio used to scale the map's latitude (vertical) distance to the corresponding pixel dimensions in the image.
• angle: The initial heading direction of the drone at the start of the route. This is expressed in degrees, with 0° being East, 90° being North, 180° being West, and 270° being South.
• destination: The final destination of the trajectory, represented as a coordinate or location within the map.
• attention_list: A list of human-annotated areas that were the focus during data collection. Each area is represented by a coordinate (latitude, longitude) and a radius (in meters) defining a circular region of interest.
• gt_path_corners: A list of four corner points that represent the boundary of each view area along the route. The corners are listed in the following order: forward left corner, forward right corner, backward right corner, and backward left corner. This provides the ground truth view area at each step of the navigation.