FOVEA-4K: Foveated Object Vision with Event-based Attention

FOVEA-4K is an object-centric active-vision dataset generated from 3D assets from the Toys4K dataset.

Rather than representing objects as independent static images, FOVEA-4K generates sequential observations through a closed sensorimotor loop that combines simulated fixational eye movements, event-camera sensing, bottom-up proto-object attention, saccadic re-centering, and foveated visual representations.

Each sequence preserves both the sensory observations and the saccadic displacement linking consecutive fixations. During each fixation, simulated drift and microsaccades generate event activity from an otherwise static object. The accumulated events drive the bottom-up proto-object saliency mechanism, whose output is converted into a spatial sampling distribution used to select the subsequent fixation location.

After event-observability quality control, 3,416 of the 4,000 candidate objects are retained.

DVS stream	Proto-object attention

Dataset contents

For each object and rendered view, FOVEA-4K provides temporally aligned:

• RGB representations
  • crops
  • foveated black-background views
  • foveated multi-scale blurred views
• luminance representations
  • crops
  • foveated black-background views
  • foveated multi-scale blurred views
• event-based representations
  • event crops
  • foveated event observations
  • raw event data
• object masks
• saccadic displacement coordinates

Each retained view contains 126 temporally ordered foveations.

Dataset structure

FOVEA-4K/
├── dataset_config.json
├── example/
│   ├── attention_debug.gif
│   └── render_events.gif
│
└── sweep_00000/
    ├── metadata_all.csv
    ├── metadata.csv
    ├── object_event_stats.csv
    │
    ├── obj_00001/
    │   ├── rot_el+30.0_az+315.0/
    │   │   ├── saccades_dxdy.dat
    │   │   ├── masks.zip
    │   │   │
    │   │   ├── RGB/
    │   │   │   ├── crops.zip
    │   │   │   ├── fov_black.zip
    │   │   │   └── fov_blur.zip
    │   │   │
    │   │   ├── LUM/
    │   │   │   ├── crops.zip
    │   │   │   ├── fov_black.zip
    │   │   │   └── fov_blur.zip
    │   │   │
    │   │   └── EVENTS/
    │   │       ├── crops.zip
    │   │       └── fov_events.zip
    │   │
    │   └── ...
    │
    └── ...

Each ZIP archive contains the complete temporal sequence for a single representation of a given object view. This organization reduces the number of individual files while preserving selective download by modality.

Dataset size

The current FOVEA-4K release occupies approximately 34.3 GB on disk (36,906,950,656 bytes) before archive-level reorganization.

Source data and licensing

The FOVEA-4K generation code is distributed separately under the license specified in the corresponding source-code repository. Repository: TBD

@misc{rodriguez_garcia_2025_15802809,
  author       = {Rodríguez García, Alejandro and Ghosh, Anindya and Ramaswamy, Srikanth and D'Angelo, Giulia},
  title        = {Object perception through visual attention},
  month        = jul,
  year         = 2025,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.15802809},
  url          = {https://doi.org/10.5281/zenodo.15802809},
}

The Toys4K assets are distributed under mixed Creative Commons licenses. Depending on the individual asset, licenses may include CC0, CC BY-SA, or CC BY-NC-SA. Users should therefore verify the license associated with each source asset and ensure that their intended use complies with the applicable terms. In particular, use of the complete dataset should be considered with respect to the most restrictive licenses represented among the source assets. Attribution and asset-specific licensing information for Toys4K is available from the original dataset resources: Stojanov, S., Thai, A., & Rehg, J. M. (2021). Using shape to categorize: Low-shot learning with an explicit shape bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Toys4K dataset page: https://rehg.org/publication/dataset2/