Angeliki Pantazi

Lightweight vision models can outperform massive baselines simply by learning where to look—boosting detection accuracy by up to 20% AP via task-guided, coarse-to-fine active glimpses instead of brute-force downsampling.