Towards robust person detection: evaluating model performance in objectifying and visually constrained scenes

This paper investigates the performance of person detection models—specifically YOLOv11 and RTMPose—under varying and objectifying conditions in cinematic content. Using the MS COCO and MovieGraphs datasets, we analyze model performance across different dimensions, including visible body parts and scene type (objectifying versus non-objectifying). Our methodology comprises systematic data preparation, targeted detection, objectification analysis, and model optimization. The results highlight key trends and limitations in person detection within complex film scenes, emphasizing the importance of detailed analysis in such environments. These findings inform future work, particularly the potential integration of the MovieNet dataset or additional multimodal strategies to enhance model robustness and fairness.