Seangmin Lee

Jiye Won, Seangmin Lee, Soon Ki Jung

Video Capsule Endoscopy (VCE) poses a challenging multi-label temporal classification problem, requiring simultaneous localization of 8 anatomical regions and detection of 9 pathological findings across tens of thousands of frames. We present GALAR-TemporalNet v2, a hierarchical temporal model that addresses three core challenges: extreme class imbalance, long-range temporal dependencies, and pathology--anatomy entanglement. Our architecture combines windowed self-attention for local modeling, a Dual-Graph GCN for global frame relationships, and Bidirectional Mamba for selective boundary context encoding. A novel anatomy prototype residual pathway decouples pathological deviation signals from normal organ appearance, and a frame-level GCN skip connection stabilizes training of visually confusable rare classes. The competition version, GALAR-TemporalNet, achieved an overall mAP@0.5 of 0.2644 and mAP@0.95 of 0.2353 on the RARE-VISION test set. Following the competition, the redesigned GALAR-TemporalNet v2 -- incorporating a restructured pathology branch, refined loss functions, and extended post-processing -- improved these results to mAP@0.5 of 0.3409 and mAP@0.95 of 0.3333.

Jinhyeong Park, Shaheryar Muhammad, Seangmin Lee et al.

Current face de-identification methods that replace identifiable cues in the face region with other sacrifices utilities contributing to realism, such as age and gender. To retrieve the damaged realism, we present FLUID (Face de-identification in the Latent space via Utility-preserving Identity Displacement), a single-input face de-identification framework that directly replaces identity features in the latent space of a pretrained diffusion model without affecting the model's weights. We reinterpret face de-identification as an image editing task in the latent h-space of a pretrained unconditional diffusion model. Our framework estimates identity-editing directions through optimization guided by loss functions that encourage attribute preservation while suppressing identity signals. We further introduce both linear and geodesic (tangent-based) editing schemes to effectively navigate the latent manifold. Experiments on CelebA-HQ and FFHQ show that FLUID achieves a superior balance between identity suppression and attribute preservation, outperforming existing de-identification approaches in both qualitative and quantitative evaluations.