Search papers, labs, and topics across Lattice.
This paper introduces ProtoBlend, a novel framework for video dataset distillation that bypasses gradient-based optimization by employing a construction-based approach. The method effectively selects informative temporal segments, allocates prototypes based on intra-class variations, and blends these prototypes with in-cluster anchors to enhance the information content of the distilled videos. Experimental results across four action-recognition benchmarks reveal that ProtoBlend achieves a competitive accuracy-efficiency trade-off, significantly improving the distillation process without the computational burden of iterative optimization.
By eliminating the need for iterative optimization, ProtoBlend achieves high-quality video dataset distillation while dramatically reducing computational costs.
Video dataset distillation aims to compress a large video dataset into a compact surrogate set that preserves its training utility. Most existing approaches synthesize condensed videos through iterative optimization, whose cost is amplified by the temporal dimension. Rather than further reducing the number of optimized variables, we investigate whether effective distilled videos can be constructed without gradient-based optimization of the stored videos. Such a construction-based approach must address three challenges: selecting informative temporal segments, covering diverse intra-class variations under a limited videos-per-class budget, and increasing the information carried by each stored sample. To this end, we propose ProtoBlend, an efficient select-allocate-blend framework. First, teacher-guided temporal clip selection retains a high-confidence segment from each source video. Second, cluster-guided prototype allocation partitions the selected clips in the teacher feature space and assigns one distilled slot to each intra-class cluster. Third, each prototype is blended with an in-cluster anchor, while their teacher predictions are combined using the same coefficient to provide mixture-source supervision. Experiments on four trimmed action-recognition benchmarks demonstrate that ProtoBlend achieves a competitive accuracy-efficiency trade-off without iterative optimization of the distilled videos.
