← SEEC documentation

SEEC dataset documentation

Purpose and composition

SEEC (Standardized External-Eye Corpus) is a large-scale external-eye/periocular image corpus developed for representation learning and benchmarking tasks related to the external eye. The corpus contains 4,352,692 unilateral 224x224 periocular crops derived from 14 public face-image datasets, together with a 134,969-image high-resolution subset at 512x512. Counts refer to images rather than unique individuals.

The accompanying benchmark includes the following resources:

These resources serve different experimental roles and should not be summed as independent additions to the corpus.

Source datasets and construction

SEEC was constructed from CelebA, UTKFace, FFHQ, VGGFace2, Tufts Face Database, Chicago Face Database, CFD-MR, CFD-India, IMDB, WIKI, UMDFaces, Face Images with Marked Landmark Points, FEI Face Database, and MORPH. See canonical source dataset documentation.

Corpus construction included deduplication, face detection and alignment, manual periocular quality annotation, dataset-specific CNN filtering, bilateral periocular cropping, left/right splitting, and resizing. Released reconstruction manifests record the transformations and dataset membership required to reconstruct the public portions of SEEC from source data. Source images are not redistributed by SEEC.

Benchmark use

The SEEC benchmark evaluates learned representations across four settings: embedding geometry under distribution shift, anatomical landmark recovery, cross-dataset anatomical transfer, and disease classification using frozen probes and short supervised fine-tuning.

Public benchmark resources are intended to support representation-learning research, model comparison, and reproducibility studies involving external-eye and periocular images.

Access and documentation

The SEEC project page is available at PROJECT_URL (TODO: replace after deployment). The dataset reconstruction toolkit is maintained at https://github.com/monkeygobah/seec-dataset, and the benchmark code and fixed evaluation resources are maintained at https://github.com/monkeygobah/seec-benchmark. Reconstruction manifests and trained checkpoints are hosted on Google Drive and are linked from the project page. The restricted clinical resources used for the Clinic and Disease evaluations are not publicly distributed.

Intended use, limitations, and human data

SEEC is intended to support research on external-eye and periocular representation learning and benchmarking. Its primary use is as a standardized resource for studying whether models trained on large-scale external-eye imagery learn anatomical and clinically relevant representations that transfer across datasets and tasks.

SEEC is derived from previously released face-image datasets rather than from direct collection of new public-subject data by the SEEC authors. Collection procedures, consent mechanisms, demographic composition, and original terms therefore differ across source datasets and should be interpreted according to the documentation provided by the respective dataset creators.

Because SEEC combines images from heterogeneous public face datasets, its scale should not be interpreted as evidence of demographic or clinical representativeness. Periocular cropping removes much of the surrounding facial context but does not guarantee anonymity or eliminate the possibility of identification. SEEC does not provide a harmonized demographic characterization or validated subject-level metadata across the combined corpus.

The accompanying clinical evaluations are intended to assess representation quality rather than establish deployment-ready clinical performance. The Clinic and Disease datasets are restricted and are not publicly distributed. SEEC is not intended for face recognition, identity verification, or direct clinical deployment.