Master's Thesis: Detection and Segmentation of Suggestive Clothing

Background/Motivation: Models that can recognise human skin, body parts, or scenes are often used to detect erotic and pornographic material. With the help of appropriate datasets [1], classification and object detection models can be trained. However, there are also images that are obviously erotic or pornographic, but cannot be recognised by conventional methods. This applies, for example, to people in skin-tight latex or leather clothing. Existing approaches in the field of "Human Parsing" can already segment people and their clothing well. Additionally, datasets like Fashionpedia [2] exist, which include segmentation masks and labels for clothing items. Objective: The aim of this master's thesis is to investigate whether and to what extent clothing items can be used for the recognition of erotic and pornographic imagery. First, it should be researched which existing approaches are suitable for addressing the question. Gaps in existing datasets and models should be described and filled with our own data and models. Based on the developed methods, it should then be evaluated whether (1) reliable detection of erotic clothing is possible and (2) whether erotic and pornographic images can be distinguished from other categories based on the recognised clothing. In this context, different counter classes should be evaluated, such as everyday, sports, or beach images. Results: As part of this master's thesis, the following results are to be achieved: • Dataset with annotations for the detection of erotic clothing. • Implementation of new approaches for the detection and segmentation of erotic clothing items. • Classification of the detected garments. • Evaluation of the models, both in terms of object detection/segmentation and classification (pornographic/erotic/normal).