Commercials detection is very important for TV broadcast analysis. However, independent classification of video shots is very difficult because a considerable portion of individual commercial shots look like program very much. In this paper, the authors proposed a novel way to tackle this problem: to treat successive video shots dependently and improve the final classification performance by considering their temporal coherence. Following this idea, the authors discussed how to apply the majority-based windowing and minority-based merging techniques to the training and test process of statistical classifiers. As a result, a new algorithm named Time-Constraint Boost is proposed. Simulation results show that this algorithm can improve both the training and generalization performance and lead to a promising commercials detection accuracy.