Object recognition is a versatile capability. Automatic guided tours and augmented reality are just two examples. Humans seem to do it subconsciously - unaware of the extensive processing required for it - while it is a complex task for machines. Methods based on SIFT features have proven to be robust for recognition. However, a prior detection step is required to limit confusion, caused by, e.g., scene clutter. We present an attention-guided method that offloads this to humans through eye tracking. Gaze data is used to extract candidate patches to recognize afterwards. It improves upon previous work by automatically selecting the dynamic size of said patch, instead of fixed large local region. Therefore increasing robustness and independence compared to fixed window size technique.