Projects & Datasets

Augmentation for Weather-Adverse Robustness Evaluation — a comprehensive benchmark framework for evaluating semantic segmentation under adverse weather conditions. Comprises three complementary pipelines: SWIFT (augmentation), PRISM (generation), and PROVE (evaluation) with 25 augmentation strategies.

Weather-balanced augmented dataset containing 71,400 images generated from Cityscapes using 25 distinct augmentation strategies. Designed for comprehensive evaluation of segmentation model robustness across diverse weather and lighting conditions.

Sky and cloud segmentation in equirectangular panoramic images. Dataset of 600 annotated 360° images with a neural network pipeline adapted for the geometric distortions inherent in equirectangular projections.

Neural network for sky and cloud segmentation from standard perspective images. Includes the SkyCloud dataset, trained models, and a complete training/evaluation pipeline.

Multi-scale semantic segmentation of universal outdoor scenes. The OUTSIDE15k dataset provides 15,000 images with pixel-level annotations across 24 outdoor scene classes, paired with the outsideNet architecture.

Neural-network-based traffic sign recognition in 360° panoramic images for semi-automatic road maintenance inventory. Dataset of 2,500 panoramic images with annotations for 81 traffic sign classes.