EVA: Human-Aligned Hard-Attention Vision

EVA is a hard-attention active vision model designed to bridge recognition performance and human-like scanpath behavior. The model observes images through sequential local glimpses and evaluates both classification accuracy and gaze-like visual exploration.

The project investigates whether active vision models can become more human-centered without relying only on full-image processing. Evaluation includes classification accuracy, scanpath similarity, fixation behavior, spatial exploration, and gaze alignment metrics such as DTW, ScanMatch, NSS, AUC, and GCS.

Code: pengcheng-pan97/EVA

Keywords: active vision, hard attention, human-like scanpaths, interpretable AI, gaze alignment, PyTorch.