Abstract:
Attire or clothing is a very important part of any one’s life. The way a person dresses defines his personality. In the age of data, it is very important to correctly detect and classify attire for various reasons. Attire detection can be used to find preferences of different people in different regions. This information is very valuable for fashion designers. Real-time clothing recognition can be useful in the surveillance context, where information about individual’s clothes can be used to identify crime suspects.
With the rising of autonomous cars, intelligent video surveillance, facial sensing and various people numbering applications, accelerated and precise target detection schemes are in high demand. This gives rise to the need of not only detecting the objects in videos or images but also creating the need to create a bounding box around them. The real time applications of this need make this a very hard task than its computer vision predecessor, image classification. Some of the challenges for this task include classification of clothing that shares similar characteristics and also detecting where the clothing is in the image. Some of the challenges for this task include classification of clothing that shares similar characteristics and also detecting where the clothing is in the image.
In our work we have focused on four tasks. (1) developing a unique dataset to detect attire; (2)multiclass classification of attire; (3) attire object detection; and (4) designing a novelarchitecture to perform simultaneous classification and localization of attire. In images, same clothes can look very easily different due to the position of the person that is wearing the clothes. Certain types of clothing can be small, and clothing types can look very different depending on aspect ratio of the input images. We have found that the size of the attire and quantity of the labeled classes has a huge effect on the performance of our Algorithm.
