Evaluating Scene Recognizability of a View


Abstract:
It is important to understand which view is better for recognizing and reconstructing a scene for many robotic applications, especially in a cluttered environment, where objects interact and may occlude one another in all views. In this paper, we introduce a novel, learning-based approach to evaluate scene recognizability from a view based on the quality and quantity of recognized objects, the recognition uncertainty, and the background recognizability, rather than the visibility. Our study shows that increasing visibility does not guarantee better recognizability of objects. The introduced view evaluator can
better characterize which view is more useful for the purpose of autonomous object recognition and scene reconstruction. The approach is validated through experiments, and the effects ofmany factors to scene recognizability are discussed based on the experimental results.

Some Experiment Results:

Fig. 1 The images are in the order ranked by our trained view evaluator from the best to the worst, and the ranks are 1, 2, 2, 2, 2, 6, 6, 8.

Fig. 2 The images are in the order ranked by our trained view evaluator from the best to the worst, and the ranks are 1, 1, 3, 4, 4, 6, 7, 8.
            
More details can be found in this paper:
• Zhou Teng and Jing Xiao, “A Learning-based Approach for Evaluating Scene Recognizability of a View”, in IEEE International Conference on Robotics and Automation (ICRA), Seattle, Washington, USA, 2015. (PDF)