ProtoBlend: A New Method for Efficient Video Dataset Distillation
Researchers have introduced ProtoBlend, a novel framework for video dataset distillation that avoids gradient-based optimization of stored videos. The method addresses three key challenges: selecting informative temporal segments, covering intra-class variations under a limited videos-per-class budget, and increasing information per stored sample. ProtoBlend operates in three stages: teacher-guided temporal clip selection retains high-confidence segments from source videos; cluster-guided prototype allocation partitions the data; and blending constructs the distilled videos. This approach promises to significantly reduce the computational cost of video dataset distillation, which is typically amplified by the temporal dimension. The paper is available on arXiv under the identifier 2608.03269.
Key facts
- ProtoBlend is a framework for video dataset distillation.
- It avoids gradient-based optimization of stored videos.
- It uses teacher-guided temporal clip selection.
- It employs cluster-guided prototype allocation.
- It addresses three challenges: temporal segment selection, intra-class variation coverage, and information density.
- The method is described in an arXiv paper with ID 2608.03269.
- The paper was announced as a cross-type submission.
- The approach aims to reduce computational cost in video dataset distillation.
Entities
Institutions
- arXiv