Active learning for sketch recognition and active scene learning

Abstract

Çizim, fikirlerin ifade edilmesi ve paylaşılması için doğal ve etkili bir araçtır.Bu nitelikler çizimin kalem tabanlı sistemler için yeni bir etkileşim kipi olmasınısağlıyor. Çizim tabanlı arayüzlerin kullanılabilirliği başarı çizim tanıma sistemlerininvarlığına dayanır ki bu da çok sayıda etiketlenmiş verinin model eğitimi içinkullanılmasını gerektirir. Ne yazık ki, çizim verisini etiketlemek zaman alıcı ve masraflıdır. Çünkü etiketleme için insanların katılımı gereklidir. Bu çalışmada, hedeflenentanıma başarısı için gerekli manuel etiketleme yükünün azaltılmasında

Sketching is a natural and effective means for expressing and sharing ideas. Thesequalities have made sketching an emerging interaction modality in pen-based systems.Sketch-based interfaces rely on the availability of accurate sketch recognition engines,which in turn require large amounts of labeled data for training. Unfortunately, labelingsketch data is time consuming and expensive, because it requires the involvementof human annotators. We demonstrate the utility of the active learning technologyin reducing the amount of manual annotation required to achieve target recognitionaccuracy.The first part of our work presents the first comprehensive study on the use ofactive learning for isolated sketch recognition. We present results from an extensiveanalysis which shows that the utility of active learning depends on a number of practicalfactors that require careful consideration. These factors include the choices ofbatch selection strategies, informativeness measures, seed set size, and domain-specificfactors such as feature representation and the choice of database. Since active learningcommunity lacks such factor based analysis, our empirical analysis is examplary.Our results imply that the Margin-based informativeness measure consistently outperformsother measures. We also show that the use of active learning brings definitiveadvantages in challenging databases when accompanied with powerful feature representations.The second part of our work deals with active learning on sketches containing morethan one object, the so-called /scenes