View centered video-based object recognition for lightweight devices

摘要

Video-based object recognition faces the problem of multi-view object variance, noisy conditions, and limited computational resources. In our previous work, we introduced a multi-view recognition approach with a compact global image descriptor coupled with orientation sensor data. Since our purpose is to run all computations in a handheld device, contrary to more intensive deep learning approaches, now we investigate the efficiency of our approach using a full representation image model with KD-Tree indexing. Experimental results show the effectiveness of our approach through three databases using noisy images.