Method for evaluating classification models for video stream analysis

In recent years, the relevance of computer vision technologies has significantly increased, especially in tasks related to video stream analysis, such as fatigue detection, face spoofing prevention, and gesture recognition. In well-known approaches for evaluating models used in these tasks, the ROC AUC metric is widely used. However, its application, based on frame-by-frame analysis, has certain limitations. These limitations are related to the instability of results due to inconsistent model outputs between frames and the lack of consideration for temporal dependencies in the data, which reduces the accuracy of the evaluation in real-world streaming video usage. The proposed methodology addresses these issues by constructing ROC curves for the video as a whole, rather than for individual frames. This process is implemented through the analysis of model outputs on each frame, allowing for more reliable differentiation between positive and negative examples at the level of the entire video segment. This approach enables a more realistic assessment of the trade-off between false positives and recall, and it also improves the sensitivity of models by selecting more relevant classification thresholds. Thus, the new methodology offers a more relevant evaluation of models for streaming video processing, ensuring more accurate event recognition and their contextual interpretation, which is especially important in critically significant applications.