Aesthetic Assessment of Movie Still Frame for Various Field of Views

As a highly influential form of visual art, films often utilize varying shot scales to communicate specific content and emotions. Given that manual annotation of each attribute in large datasets is impractical, our research aims to automate the aesthetic evaluation of movie frames. We constructed a dataset of 24,500 movie still frames, which were annotated with overall aesthetic scores by film experts. We then developed aesthetic evaluation models tailored to different scenes, calculating attribute scores, attribute contributions, and predicted overall scores for each frame. Ablation experiments validated the effectiveness of our feature design. Subjective assessments by film experts also confirmed the reliability of our model's computational outcomes, endorsing its capability to quantitatively and aesthetically evaluate movie frames.