A personalized active method for 3D shape classification

Abstract

To efficiently and flexibly classify 3D shape in a user-adaptive way, this paper proposes a novel interactive system by incorporating active learning, online learning and user intervention. Given a shape collection, our system iteratively alternates the interactive annotation and verification until the labels of all the shapes are confirmed explicitly by the users. The main advantage is that it provides faster interactive classification rates than alternative approaches. Our system achieves this goal by a unified active learning algorithm that selects the shapes to be annotated or verified for the next iteration. The shape selection step is solved by maximizing a probability model for simulating the time cost of human input during manual intervention. The selected shapes are then pushed to the users along with their category predictions, which are generated by a proposed weighted blending framework. After manually confirming or refining the automatic predictions, we use an extended soft confidence-weighted learning method to update the classifier incrementally for the subsequent active selection and shape classification in turn. Experimental results demonstrated the effectiveness of the proposed method.

Personalized Image Classification by Semantic Embedding and Active Learning

“…Yi et al [ 22 ] propose an active framework for annotating massive 3D shape datasets. Song et al [ 23 , 24 ] further realize iterative 3D shape classification by combining active learning and online learning. Our method is inspired by these work, and extending them to support personalized image classification.…”

“…In the area of 3D shape classification, the state-of-the-art MVCNN [ 25 ] only requires 128-dimensional vector to obtain the classification accuracy more than in ModelNet40 [ 26 ]. With its low dimension and high discrimination, this yields very good performance for interactive 3D shape classification [ 23 , 24 ]. However, for image classification, as the variation between images is much larger [ 27 ], the dimension of the feature layer is above 2000 in the typical network model.…”