AAA is an open-source dataset for vision-and-language research and accessibility. It provides four Angles of View (AoV) to enable controlled evaluation across camera geometries: 70°, 110°, 180° (fisheye), and 360° (equirectangular).
To support assistive applications for visually impaired users, AAA also includes a targeted Q&A benchmark organized into four categories:
Together, these components enable fair comparison of models across AoVs while evaluating practical, assistive questions that matter in real-world settings.
We do not redistribute GND. Please download from the official site: GND dataset.
git clone https://github.com/Pooh1101/All-Around-Assistance-AAA-Dataset.git
cd All-Around-Assistance-AAA-Dataset
Ensure the structure is as follows:
frames_folder_root/
├── AU/
│ ├── S01/
│ │ ├── frame0000
│ │ ├── frame0001
│ │ └── ...
│ ├── S02/
│ │ └── ...
│ └── ...
│
└── CUA/
└── ...
# build views (70 / 110 / 180 / 360)
python make_views_min.py </path/to/frames_folder_root> </path/to/output_folder>
#run Mask2Former
python mask2former_build_masks.py </path/to/output_folder>/frames --model facebook/mask2former-swin-large-coco-panoptic
19 commits
Python
100.0%
AAA is an open-source dataset for vision-and-language research and accessibility. It provides four Angles of View (AoV) to enable controlled evaluation across camera geometries: 70°, 110°, 180° (fisheye), and 360° (equirectangular).
To support assistive applications for visually impaired users, AAA also includes a targeted Q&A benchmark organized into four categories:
Together, these components enable fair comparison of models across AoVs while evaluating practical, assistive questions that matter in real-world settings.
We do not redistribute GND. Please download from the official site: GND dataset.
git clone https://github.com/Pooh1101/All-Around-Assistance-AAA-Dataset.git
cd All-Around-Assistance-AAA-Dataset
Ensure the structure is as follows:
frames_folder_root/
├── AU/
│ ├── S01/
│ │ ├── frame0000
│ │ ├── frame0001
│ │ └── ...
│ ├── S02/
│ │ └── ...
│ └── ...
│
└── CUA/
└── ...
# build views (70 / 110 / 180 / 360)
python make_views_min.py </path/to/frames_folder_root> </path/to/output_folder>
#run Mask2Former
python mask2former_build_masks.py </path/to/output_folder>/frames --model facebook/mask2former-swin-large-coco-panoptic
19 commits
Python
100.0%