External-eye photos get SEEC, a 4.35-million-image corpus: SEEC-pretrained periocular encoders beat DINOv2 and MAE on cross-dataset landmark transfer
Synopsis
External-eye photographs had no large-scale standardized resource, since existing face datasets were built for identity recognition or image generation and dedicated iris and periocular datasets are far smaller than modern pretraining corpora. SEEC turns fourteen public face datasets into 4,352,692 standardized unilateral periocular crops and pairs them with four evaluation axes: embedding geometry, anatomical landmarks, cross-dataset transfer and disease classification. On cross-dataset landmark transfer, the best SEEC-pretrained ViT encoder beats DINOv2 and MAE on mean L2, mean MAE, iris centroid, medial canthus and lateral canthus error, despite pretraining at much smaller scale.
Interpretation
External-eye photography records clinically relevant anatomy of the eyelids, palpebral fissures, canthi, iris, sclera and brows at low cost, yet no resource combined scale, standardized periocular anatomy and an evaluation structure for studying reusable external-eye representations. Existing face datasets were built for identity recognition, facial attributes or image generation, while dedicated iris and periocular datasets target biometric recognition or specialized eye-analysis tasks and are far smaller than contemporary pretraining corpora. SEEC starts from fourteen public face datasets and applies deduplication, face detection and landmark alignment, manual periocular annotation, dataset-specific CNN filtering, bilateral cropping and left/right splitting, yielding 4,352,692 unilateral 224×224 periocular images plus a 134,969-image 512×512 high-resolution subset.
Evaluation separates the representation level, which asks how the embedding space is occupied across in-distribution, high-resolution and clinically shifted distributions, from the functional level, which asks whether frozen representations recover anatomical landmarks, transfer across datasets and support disease classification. A single downstream endpoint cannot answer embedding geometry, anatomical fidelity and clinical separability at once, so the benchmark computes normalized effective rank on Holdout, Open-HR and Clinic and runs functional probes on LM-Celeb, LM-CFD and Disease. The two levels give dissociable conclusions: LeJEPA reaches the highest normalized effective rank for both ResNet-101 and ViT-B/16 yet does not dominate every anatomical or disease probe, and all objectives lose effective rank under distribution shift, with the largest drops on Clinic.
In cross-dataset landmark transfer, the best SEEC-pretrained ViT-B/16 encoders outperform both DINOv2 and MAE on mean L2, mean MAE, iris centroid, medial canthus and lateral canthus error. Those two general-purpose baselines were pretrained on LVD-142M and ImageNet-1K respectively, far larger than the controlled SEEC pretraining subsets. The transfer setting trains a lightweight landmark regression probe on LM-Celeb and evaluates on LM-CFD without retraining; randomly initialized ViT-B/16 InfoNCE achieves the lowest mean L2 and mean MAE of all evaluated encoders, and randomly initialized ViT-B/16 LeJEPA achieves the lowest medial and lateral canthus errors.
On frozen linear-probe five-class disease classification, SEEC-pretrained ViT-B/16 models are strongest overall, with InfoNCE achieving the highest accuracy, recall and F1 and VICReg second. DINOv2 matches LeJEPA in accuracy and F1 on the frozen probe but stays below the best SEEC-pretrained ViT-B/16 models, while MAE and all ResNet-101 models perform lower overall. The disease dataset contains 633 images across thyroid eye disease, chalazion, Treacher Collins syndrome, ptosis and Crouzon/Apert/Pfeiffer syndromes, split 80/20 and grouped at the source-image level to prevent leakage.
Perspective
Evaluation covers two backbones, ResNet-101 and ViT-B/16, three self-supervised objectives, InfoNCE, VICReg and LeJEPA, and pretraining subsets of 10,000, 100,000 and 1,000,000 images, with downstream evaluation using frozen backbones by default. The anatomical probe covers only nine iris, scleral, canthal and eyelid landmarks derived from segmentation masks, across the LM-Celeb and LM-CFD datasets. Clinical evaluation uses 35,082 unlabeled Clinic images for geometry analysis and 633 five-class disease images for classification, neither of which is publicly released due to institutional and ethical restrictions.
Disease classification is grouped at the source-image level rather than by confirmed patient-level separation because patient-level identifiers were unavailable. After five epochs of supervised adaptation, ResNet-101 loses performance on disease classification, DINOv2 is strong on the frozen probe but degrades substantially under fine-tuning, and t-SNE shows overlap between some disease classes. Controlled pretraining used only a two-view augmentation setup, and the authors propose richer local/global multi-view training, longer pretraining schedules and larger models as next steps.
