Curriculum Learning of Visual Attribute Clusters for Multi-Task Classification

摘要: Visual attributes, from simple objects (e.g., backpacks, hats) to soft-biometrics gender, height, clothing) have proven be a powerful representational approach for many applications such as image description and human identification. In this paper, we introduce novel method combine the advantages of both multi-task curriculum learning in visual attribute classification framework. Individual tasks are grouped after performing hierarchical clustering based on their correlation. The clusters learned setup by transferring knowledge between clusters. process within each cluster is performed setup. By leveraging acquired knowledge, speed-up improve performance. We demonstrate effectiveness our via ablation studies detailed analysis covariates, variety publicly available datasets humans standing with full-body visible. Extensive experimentation has that proposed boosts performance 4% 10%.