Summary of Advancing Compressed Video Action Recognition Through Progressive Knowledge Distillation, by Efstathia Soufleri et al.
Advancing Compressed Video Action Recognition through Progressive Knowledge Distillation
by Efstathia Soufleri, Deepak Ravikumar, Kaushik Roy
First submitted to arxiv on: 2 Jul 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper proposes a novel approach to compressed video action recognition by leveraging the different modalities in compressed videos, including motion vectors, residuals, and intra-frames. Three neural networks are deployed, each dedicated to processing one modality, and the results indicate that the network processing intra-frames tends to converge to a flatter minimum than the others. The authors propose Progressive Knowledge Distillation (PKD), a technique that incrementally transfers knowledge across modalities to achieve flatter minima, which are generally associated with better generalization. PKD involves attaching early exits (Internal Classifiers – ICs) to the three networks and distilling knowledge starting from the motion vector network, followed by the residual, and finally, the intra-frame network. The authors also propose Weighted Inference with Scaled Ensemble (WISE), which combines outputs from the ICs using learned weights, boosting accuracy during inference. Experimental results show that training the ICs with PKD improves accuracy by up to 5.87% and 11.42% on the UCF-101 and HMDB-51 datasets, respectively. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper introduces a new way to understand videos by combining information from different parts of the video. This is done by using special neural networks that are trained to look at different types of information in the video. The results show that this approach can improve accuracy by up to 5-11% on certain datasets. The authors also propose a new way to combine the outputs of these networks, which helps even more. |
Keywords
* Artificial intelligence * Boosting * Generalization * Inference * Knowledge distillation
