Publications

We reproduce and re-evaluate XFeat, a lightweight image-matching model designed for efficient feature extraction on resource-constrained hardware. Our experiments recover its strong accuracy–efficiency trade-off on standard benchmarks, while architectural ablations clarify the role of its keypoint branch and skip connection. We also evaluate robustness under out-of-distribution and cross-modal image matching, where performance declines under severe modality shifts.

We create a Slovene sarcasm-detection dataset using machine translation and large language models, then evaluate cross-lingual transfer across monolingual and multilingual Transformer models. Our ensemble reaches an F1 score of 0.765.