DICENet: Fine-Grained Recognition via Dilated Iterative Contextual Encoding

Material Recognition is an intriguing problem in Computer Vision. While traditional approaches prefer an ensemble of networks to capture essential properties such as texture, more recent approaches leverage the power of Deep Learning to design end-to-end models. We do the same, and propose Dilated Iterative Contextual Encoding Network, a novel end-to-end framework for material recognition. As a result of gathering extensive knowledge on various characteristics of materials, our approach combines different components on the base network to address specific properties, which also helps in general recognition tasks. The traditional ResNet is replaced by a Dilated Residual Network to help capture fine-grained material information. Iterative deep aggregation helps capture and fuse global homogeneous material properties across multiple resolutions and scales. To enhance the discriminatory power of the learnt latent representation, we propose gramedial loss which is intuitively applied on a texture vector space. Spatial similarity loss is applied on strategic intermediate feature maps to effectively capture local non-homogeneous texture features from a global context, crucial for the primary classification task. Extensive experiments conducted on golden material datasets such as the FMD, MINC-2500, KTH-TIPS-2b, DTD and GTOS indicate improved performances over state of the art approaches on large datasets and two small datasets, while achieving compatible accuracies on the challenging FMD. Furthermore, our architecture also performed convincingly while categorizing general indoor and object classification datasets such as MITIndoor and CalTech-101.