cywan/StreamVLN-Trajectory-Data

This repo contains the data for the paper "StreamVLN: Streaming Vision-and-Language Navigation via SlowFast Context Modeling."

Overview

The dataset consists of visual observations and annotations collected in the Matterport3D (MP3D) environment using the Habitat simulator. It combines data from several open-source Vision-and-Language Navigation (VLN) datasets.

Data Collection

Data collected in this repo are from the following open-source datasets:

• R2R-VLNCE
• RxR-VLNCE
• R2R-EnvDrop

To get actions and observations, we enable a ShortestPathFollower agent in the Habitat simulator to follow the subgoals and collect rgb observations along the path. The data is collected across the Matterport3D (MP3D) scenes.

Dataset Description

Dataset Structure

After extracting images.tar.gz, the dataset has the following structure:

StreamVLN-Trajectory-Data/      
├── R2R/
│   ├── images/
│   │   ├── 1LXtFkjw3qL_r2r_000087/
│   │   │   └── rgb/
│   │   │       ├── 000.jpg
│   │   │       ├── 001.jpg
│   │   │       └── ...
│   │   ├── 1LXtFkjw3qL_r2r_000099/
│   │   ├── 1LXtFkjw3qL_r2r_000129/
│   │   └── ...
│   └── annotations.json     
├── RxR/
│   ├── images/
│   └── annotations.json      
└── EnvDrop/
    └── annotations.json   

Contents

images/: The folder contains the rgb observations collected from Habitat simulator.

annotations.json: The file contain the navigation instructions and discrete actions sequence from Habitat Simulator for each dataset. The structure of annotation for each episode is as follows:

{
    "id": (int) Identifier for the episode,
    "video": (str) Video ID to identify the relative path to the directory which contains the episode, format: "images/{scene}_{dataset_source}_{id}",
    "instruction": (list[str]) Navigation instructions,
    "actions": (list[int]) Discrete actions sequence in Habitat simulator, 
                # 1 = MoveForward (25cm)
                # 2 = TurnLeft (15°)
                # 3 = TurnRight (15°)
                # -1 = Dummy
                # 0 = Stop (omitted in annotations)
}

Each episode in the annotations.json file corresponds to a folder in the images/ directory, where the folder name is included in the video ID. The rgb images are stored in the rgb/ subdirectory of each episode folder. Length of the actions list corresponds to the number of rgb images in the episode to ensure observation-action data pairs.

EnvDrop Dataset Note

For EnvDrop, only navigation annotations are provided due to the large number of episodes. To obtain RGB observations, you can replay the annotated actions using the Habitat simulator. Below is a example demonstrating how to replay stored actions and get RGB frames:

1. Before proceeding, you need to modify the configuration file to specify the path to the EnvDrop episodes. Please overwrite the habitat.dataset.data_path in config/vln_r2r.yaml:

   habitat:
       ...
       dataset:
           ...
           data_path: data/datasets/envdrop/envdrop.json.gz
   

2. Run the code below to save RGB images.

   import os
   import json
   import habitat
   
   from habitat_baselines.config.default import get_config
   from habitat.tasks.nav.shortest_path_follower import ShortestPathFollower
   from streamvln.habitat_extensions import measures
   
   CONFIG_PATH = "config/vln_r2r.yaml"  # Path to the Habitat config file
   ANNOT_PATH = "data/trajectory_data/EnvDrop/annotations.json"  # Path to the annotations file
   GOAL_RADIUS = 0.25  # Radius for the goal in meters. not used if get actions from annotations
   
   env = habitat.Env(config=get_config(CONFIG_PATH))
   annotations = json.load(open(ANNOT_PATH, "r"))
   
   for episode in env.episodes:
       env.current_episode = episode
       agent = ShortestPathFollower(sim=env.sim, goal_radius=GOAL_RADIUS, return_one_hot=False)
       observation = env.reset()
   
       annotation = next(annot for annot in annotations if annot["id"] == int(episode.episode_id))  # Get annotation for current episode
       reference_actions = annotation["actions"][1:] + [0]  # Pop the dummy action at the beginning and add stop action at the end
       step_id = 0  # Initialize step ID
   
       while not env.episode_over:
           rgb = observation["rgb"]  # Get the current rgb observation
           
           # TODO: Save RGB frame (customize as needed)
           # --------------------------------------------------------
           import PIL.Image as Image
           video_id = annotation["video"]  # Get the video ID from the annotation
           rgb_dir = f"data/trajectory_data/EnvDrop/{video_id}/rgb"
           os.makedirs(rgb_dir, exist_ok=True)
           Image.fromarray(rgb).convert("RGB").save(os.path.join(rgb_dir, f"{step_id:03d}.jpg"))
           # --------------------------------------------------------
   
           action = reference_actions.pop(0)  # Get next action from our annotation
           observation = env.step(action)  # Update observation
           step_id += 1
   
   env.close()