Video analysis has a long-standing problem that remains unsolved: how to deal with the segmentation of video into different scenes. This matter had declined in interest, but the authors of this paper bring it back into consideration in response to the rise of video authoring on YouTube, Facebook, and the like. Currently, millions of videos are uploaded daily, and each one has to be processed, and its content classified, to be accessible to viewers. One of the first stages of classification involves the segmentation of a video clip into the different scenes that compose it.

Two main technologies have led the development of scene video segmentation: machine learning methods and multimodal techniques. However, the authors of this paper explore a third approach, based on visual coherence. This technique processes a video to extract visual features, such as the histogram, contours, movements, and so on, and computes their coherence from different frames.

The proposed method uses a visual coherence model based on histogram similarity and motion dynamics. To obtain the visual coherence, the authors extract normalized histograms based on hue, saturation, and value (HSV) (as visual features) and optical flow (as motion features). Shots are determined using the absolute difference between a histogram and the interjection of the histograms of consecutive frames. After that, the scene is segmented with a serial procedure: key-frame selection, shot coherence calculation, removal of similar adjacent scenes, comparison of motion similarity, and removal of highly similar scenes.

This procedure achieves quite high precision, accuracy, and F1 measure in scene segmentation. The authors compared the proposed method in experiments with two state-of-the-art methods. Those experiments show that, in some cases, the proposed method behaves better than either of the other two techniques.

This paper describes every step in great detail, which should enable a reader to replicate the proposed method. The authors assert that the proposed method has lower computational costs. However, they do not provide any experimental results to support that statement.

In conclusion, I found this to be an interesting method, which provides good results (although not extraordinary ones) and promises to reduce computational costs.