Pupil-DLC: 21,909 Manually Annotated Frames for Pupil Dynamics Tracking

Parsa Seyfourian, Lydia C. Marks, Leslie D. Claar, Yasmeen Nahas, Miles Keating, Christof Koch & Irene Rembado created this dataset of 21,909 manually annotated frames extracted from 145 videos. The frames were used to train the General Model for the Pupil-DLC deep learning pipeline. The dataset was published on figshare under a CC-BY-4.0 license on June 4, 2026.

Use Cases

• Training deep learning models for marker-less pupil tracking based on the manually annotated frames.
• Benchmarking the performance of pupil detection algorithms across conscious and unconscious states.
• Developing scalable computer vision pipelines for behavioral neuroscience experiments.

Strengths

• Contains 21,909 manually annotated frames, providing a substantial training resource.
• Derived from 145 source videos, suggesting diversity in the underlying data.
• Released under a permissive CC-BY-4.0 license for open use and sharing.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for specific model architectures.

Provenance

Source

Irene Rembado (author) via figshare.

Collection Method

Manually annotated frames extracted from videos for training a deep learning model.

Freshness

Last updated 2026-06-04 22:38:29; freshness should be verified.