README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: mask
      dtype: image
  splits:
    - name: train
      num_bytes: 840243388.6753247
      num_examples: 207
    - name: test
      num_bytes: 97419523.32467532
      num_examples: 24
  download_size: 919773656
  dataset_size: 937662912
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
license: mit
task_categories:
  - mask-generation
tags:
  - street
  - view
  - street-view
  - '360'
pretty_name: 360° streets view with mask

📷 360 Clean Dataset

A dataset of 360° equirectangular images with corresponding binary masks that hide the typical artifacts introduced by 360° capture, such as:

• 🚗 Vehicles (cars, bikes, etc.),
• 🧍‍♂️ The person capturing the video (cyclist, pedestrian, etc.),
• 🎥 Camera equipment or shadows appearing at the bottom of the image.

🧾 Description

Each sample in the dataset contains:

• image: the original 360° equirectangular image (2:1 aspect ratio, typically 3040×1520),
• mask: a binary mask of the same resolution, where white pixels (255) indicate areas to ignore (e.g. person, vehicle), and black pixels (0) represent the usable background.

The masks were manually created.

This dataset is particularly useful for:

• 🗺️ 3D reconstruction tasks (e.g. NeRF, Gaussian Splatting),
• 🤖 Training vision models without human-related artifacts,
• 📍 Visual geolocation from clean, unobstructed environments.

📁 Data Format

{
    "image": Image,  # equirectangular 360° scene
    "mask": Image    # binary mask: 1 = ignore, 0 = keep
}

Files are matched by filename: xxx.jpg and xxx_mask.png.

🏷️ Possible Use Cases

• Object removal / Inpainting
• Semantic Segmentation
• Dynamic object filtering
• Preprocessing for 3D or geospatial vision tasks

The model Jour/sam-vit-base-equirectangular-finetuned is trained using this dataset.

🪪 License

This dataset is released under the MIT.