Sketch-specific data augmentation for freehand sketch recognition

Sketch recognition remains a significant challenge due to the limited training data and substantial intra-class variance of freehand sketches for same object. Conventional methods this task often rely on availability temporal order sketch strokes, additional cues acquired from different modalities supervised augmentation datasets with real images, which also limit applicability feasibility these in scenarios. In paper, we propose novel sketch-specific (SSDA) method that leverages quantity quality automatically. From aspect quantity, introduce Bezier pivot based deformation (BPD) strategy enrich data. Towards improvement, present mean stroke reconstruction (MSR) approach generate set types smaller variances. Both solutions are unrestricted any multi-source sketches. Furthermore, show some recent deep convolutional neural network models trained generic classes images can be better choices than most elaborate architectures designed explicitly recognition. As SSDA integrated networks, it has distinct advantage over existing methods. Our extensive experimental evaluations demonstrate proposed achieves state-of-the-art results (84.27%) TU-Berlin dataset, outperforming human performance by remarkable 11.17% increase. Finally, more experiments practical value our sketch-based image retrieval.