Fashion Learning adapts to Attributes’ Data

Aleesha Institute of Fashion Designing

The ancient Greek philosopher said that it is much easier to recognize the attributes of the concept than to understand it directly. Nowadays, his descendants employ this philosophy to improve objects understanding in each image at fine-grained level. This research direction is developed widely. These finegrained details are often called attributes. While local features are supposed to be better than global features, many handcrafted features like as SIFT, HOG, … are used to extracted visual features. However, the features are not appropriate to all attributes. In addition, to narrow semantic gap in fashion recognition and retrieval, local features are replaced by attribute vectors for each fashion image. Hence, attribute learning become a hot topic in many researches. In Computer Vision, attributes are used in face retrieval [3], fashion retrieval [2], violent’s attributes detection [17], crowd attributes [18], … Moreover, the attributes can be used as fast filters in large-scale image retrieval. In this paper, we focus on fashion attributes that contain many challenges. First, fashion attributes are always renewed with large variety in categories. Then, by the essence of fashion attributes, the imbalanced data problem often exists on real dataset. Finally, attributes in fashion show high-level semantic descriptions which share many different visual descriptions. In the traditional way, each attribute will have an individual model to learn attribute. However, the unpredicted increase in fashion attributes quantity throughout years, building a learning model for each attribute becomes ineffective.

Thanks to the development of DCNNs, we can make an attribute MTL learning model which are supposed to increase the effectiveness and efficiency of whole system. However, imbalanced data problem for fashion attributes is hard to deal. In this work, we consider learning binary fashion attributes through a deep local multi-task pre-trained CNN model based on threshold configuration in solving imbalanced data problem. To evaluate the effectiveness of this framework, we experimented on DeepFashion Attribute dataset with the pretrained Nasnet model. The experiments show that our local MTL framework achieves a better performance than MTL for fashion attribute. Experiments also show that our model considering imbalanced data achieved better performance than models that do not account for this. The contributions of our work are as follows: – We propose a local deep multi-task transfer learning framework in solving fashion attribute learning. – We propose a quite new imbalanced data problem solver for MTL which is not based on data sampling, networks refining or loss functions changing.

This work presents a new approach in solving imbalanced data problem for fashion attribute learning based on output configuration with MCC. Through extensive experiments, we show the effectiveness of our suggested method based on Nasnet transfer learning in local multi-attribute learning which paves the way to further improvement.