Camera Movement Classification in Historical Footage: A Comparative Study of Deep Video Models
Tingyu Lin, Armin Dadras, Florian Kleber, Robert Sablatnig
TL;DR
This study evaluates deep video classification models for camera movement classification (CMC) on historical footage using the HISTORIAN dataset, addressing challenges from degradation and nonstandard motion. By comparing five architectures—including C3D, I3D, R(2+1)D, TSN, and Video Swin Transformer—the work demonstrates that temporal modeling is crucial for detecting subtle, globally coherent camera motions in archival film. The Video Swin Transformer achieves the best performance at 80.25% top-1 accuracy, illustrating strong generalization despite limited data, while a traditional CAMHID baseline provides a point of comparison on a restricted subset. The findings support further exploration of multimodal inputs, reproducible benchmarking, and transfer learning to bridge domain shifts between modern and historical video data.
Abstract
Camera movement conveys spatial and narrative information essential for understanding video content. While recent camera movement classification (CMC) methods perform well on modern datasets, their generalization to historical footage remains unexplored. This paper presents the first systematic evaluation of deep video CMC models on archival film material. We summarize representative methods and datasets, highlighting differences in model design and label definitions. Five standard video classification models are assessed on the HISTORIAN dataset, which includes expert-annotated World War II footage. The best-performing model, Video Swin Transformer, achieves 80.25% accuracy, showing strong convergence despite limited training data. Our findings highlight the challenges and potential of adapting existing models to low-quality video and motivate future work combining diverse input modalities and temporal architectures.