Publication

Abstract : With rapid technological progress, remote learning through lecture videos in Massive Open Online Courses (MOOCs) has become increasingly common. However, these videos are often long and unstructured, which can negatively impact students’ attention spans in virtual learning environments, particularly in the post-pandemic context. In this work, we present ViSegEdu, a vision-only video segmentation framework that uses self-supervised Vision Transformers (DINO-v2) to capture semantic visual changes without relying on audio or transcripts. The work also makes use of Bayesian Online Change Point Detection (BOCPD) as a means to obtain Temporal Visual Semantic Changes (TVSC). This is followed by a lightweight transformer head for a fine-grained refinement. The ViSegEdu framework was evaluated on two benchmark datasets—NPTEL lectures and the LPM dataset—covering both slide-based and chalkboard-style teaching formats. The method achieves an average segmentation F1-score of 78.90% on NPTEL videos and 85.35% on LPM videos. These results highlight the effectiveness of purely visual semantic modeling for boundary detection, making ViSegEdu well-suited for large-scale video segmentation applications.