This chapter reviews related work and establishes the technical foundations for multimodal social media event detection. We first trace the evolution from Topic Detection and Tracking in structured news to modern social streams, highlighting how volume, velocity, variety, and veracity reshape assumptions, model design, and evaluation. We then introduce core principles of multimodal representation learning, including modality-specific encoders, alignment objectives, and fusion architectures ranging from early/late fusion to attention- and Transformer-based semantic integration. Next, we summarize modern deep learning approaches that expand the notion of context beyond individual posts, covering linguistic context via pre-trained language models, network context via graph neural networks, temporal context via open-world and continual learning, and discourse context via cross-task integration. Finally, we discuss why benchmark-driven evaluation can overstate deployable performance due to representation gaps, dataset biases, and modality missingness, thereby motivating the deployment-oriented datasets, protocols, and specialized frameworks developed in the subsequent chapters.