Improved Rooftop Detection in Aerial Images with Machine Learning 11/18/2018 Improved Rooftop Detection in Aerial Images with Machine Learning M. A. Maloof Georgetown University P. Langley Institute for the study of learning and expertise T. O. Binford Stanford University R. Nevatia University of southern California S. Sage Presented By Yong Li 11/18/2018
Outline Introduction The task domain Image analysis Data set description Four algorithms for supervised learning Experiments and results Comments
Introduction The flood of images Handcrafted knowledge Recent applications in business and industry hold useful lessons Machine learning approach improves the ability of image analysts to deal with the flood of images.
The Task Domain Aerial images of ground sites. A common task: detect change at a site as reflected in differences between two images (building, road, and vehicles) This paper focus on identifying buildings in satellite photographs. HOW ??
Image Analysis Building Detection and Description System (BUDDS) (Lin and Nevatia, 1998) Rooftop Generation Pixels Edge elements Edge detector lines Line feather detector U-Constructs Three-sided structure Junctions and parallel line parallelograms U-constructs and junctions
Image Analysis rooftop generation Rooftop selection Machine learning may improve this process (Hypothesis)
The Data Set Description Following Lin and Nevatia (1998) Nine continue feathers Two classes—rooftop and non-rooftop Manual labeling(interactive labeling system)
The Learning techniques Nearest Neighbor Naïve Bayes C5.0 Perceptron BUDDS Classifier (Control)
Cost-Sensitive Learning Severely skewed data Favoritism toward the majority class Naïve Predict the class i with the least expected risk C 5.0 Uses a method similar to that in CART (Breiman et al., 1984) Nearest neighbor Incorporate a cost parameter τj (0, 1). The altered distance Perceptron and BUDDS The adjusted threshold sj is the maximum value the weighted sum can take for the jth class
ROC Analysis The accuracy measure is not sufficient Separated accuracy on both classes in term of false positives and false negatives ROC (Receiver Operating Characteristic) analysis Vary some aspect of the situation (costs, the class distribution or threshold) Plot the false positive rate against the true positive rate
ROC Curve
Experiments and Results (I) Experiment I. Randomly, 60% training and 40% testing Accuracy as the measure of performance is not suitable to this problem
Experiments and Results (I)
Experiments and Results (II) Within-Image Learning Between-Image Learning Generalizing over Aspect Generalizing over Location Leaving One Image Out
Experiments and Results (II)
Experiments and Results (II)
Experiments and Results (II) Naive Naive
Experiments and Results (II)
Experiments and Results (II)
Experiments and Results (II)
Experiments and Results (II)
Experiments and Results (II)
Conclusion the results of the naïve Bayesian classifier are best one and support our main hypothesis. But not always better than BUDDS
Comments Advantages Disadvantages Comprehensively studied the use of machine learning to improve a rooftop detection process. It is possible to use machine learning methods to identify rooftops and outperform the performance of BUDDS (not strong). Disadvantages Didn’t consider time and space complexity.