CA-Market: A Partially Categorical AnnotatingApproach Based on Market1501 Dataset for Attribute Detection

Abstract

In this paper, a new partial categorical attributes dataset (CA-Market) based on images of the Market1501 dataset has been introduced, for the sake of improving the attribute detection task. Most attributes detection datasets (human appearance features detection) are not partially categorical and do not properly take into account the inner classes diversity. Increasing the diversity of inner parts (gender, head, upper-body clothes, lower-body clothes, bags, shoes, and colors) before annotating can ease the decision-making process by dividing labels into individual categories. CA-Market contains 46 binary attributes in 10 parts from head to foot and their colors which are annotated in image-level. For example, the attributes of the leg part are skirts, shorts, and pants which are carefully chosen to be categorized for a classification task. In this research, the effect of the labeling approach is studied. Hence, a common classification method is used and only datasets or baselines are changed for comparisons. Baselines are based on Omni-Scale, Resnet50, and Hydra-Plus architectures to compare the CA-Market1501 dataset with the Market1501 attribute dataset in the same setting. CA-Market demonstrates a new representation of data as a part-based format which can gain better results. This approach, without adding any extra modules, achieved a significant enhancement. For instance, accuracy in the vectorized format is over 92%, in the categorized is over 90% which shows the effectiveness of part-based attribute annotating. Also, hair, backpack, upper color, and lower color as the common attributes between Market1501-attribute and CA-Market datasets are achieved 90.26, 88.04, 94.55, and 94.18 classification accuracy which can outperform existing state-of- the-art approaches.