Designing a Face Recognition System for Event Photography

Event photography produces large collections of images that are difficult to search manually, limiting participants ability to retrieve photographs in which they appear. This work presents a face-based image retrieval system designed for this context, combining a ResNet face encoder with Euclidean distance search over a pre-indexed gallery of embeddings. The system was evaluated on SBrT2025, a real-world dataset of 1,427 images collected at a Brazilian telecommunications conference, where 81.2% of all detected faces were classified as small faces, a challenge characteristic of event photography and absent from standard face recognition benchmarks. Manual annotation of 715 images (50% of the dataset) shows that the pipeline achieves a face-presence recall of 88.5%, successfully identifying at least one face in the vast majority of images containing a visible subject. Detection accuracy varied significantly by image composition, reaching 66.8% exact match on solo images but dropping to 35.3% on group shots, where occlusion and small face scale remain the primary sources of error. These results establish a baseline for face-based retrieval in event photography and motivate future work on identity-aware evaluation and retrieval robustness under unconstrained real-world conditions.