Alptekin Temizel

Most industrial anomaly detectors collapse under simple brightness or zoom changes, losing up to 34 AUROC points, but a new approach using synthetic anomalies maintains over 92% AUROC.

LALE achieves near state-of-the-art performance in land-cover segmentation while using 4.5x fewer parameters and 17x fewer GMACs than leading models.

Synthetic data, when grounded in vision-language models for evaluation, demonstrably boosts performance in remote sensing tasks like segmentation and captioning, outperforming models trained solely on real-world data.

Forget fully adapting models: a projection-free local alignment preserves the native CLIP geometry, yielding state-of-the-art out-of-distribution detection performance.

TopoMaskV3's 3D mask head leapfrogs prior 2D methods for road topology understanding, proving that dense offset and height predictions can conquer discretization artifacts and geographic overfitting.