This thesis addresses the challenge of detailed person analysis through attribute recognition, which serves as a flexible representation of detailed personal information. Applications include autonomously operating robots and intelligent vehicles. We propose a two-phase pipeline for analysis: localization followed by attribute recognition. Contributions are made to both phases, starting with an evaluation of geometric scene constraints that enhance detector efficiency and precision. In the attribute recognition phase, we introduce the Attribute Classification Network (ACN), a deep convolutional neural network that performs surprisingly well compared to more complex methods. The ACN initially relies on externally provided weight initialization, prompting an exploration of data-dependent initialization. This leads to the development of the self-supervised PatchTask, which is experimentally explored and quantitatively evaluated in the thesis's final chapter. While the primary contributions focus on attribute recognition, aspects like the self-supervised PatchTask may have broader implications beyond this specific task.
