Pixel Distillation: Cost-Flexible Distillation Across Image Sizes and Heterogeneous Networks
IEEE Transactions on Pattern Analysis and Machine Intelligence · 2024-12 · 46 (12) · 9536-9550
DOI: 10.1109/tpami.2024.3421277 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024-12. Verfasst von Guangyu Guo und 5 weiteren Autor:innen. Die Arbeit wurde bisher 25-mal zitiert (Crossref) und nennt 84 Literaturangaben. Gefördert von: National Science and Technology Major Project, National Natural Science Foundation of China, Anhui Provincial Key R&D Programmes.
Abstract (Originalsprache)
Previous knowledge distillation (KD) methods mostly focus on compressing network architectures, which is not thorough enough in deployment as some costs like transmission bandwidth and imaging equipment are related to the image size. Therefore, we propose Pixel Distillation that extends knowledge distillation into the input level while simultaneously breaking architecture constraints. Such a scheme can achieve flexible cost control for deployment, as it allows the system to adjust both network architecture and image quality according to the overall requirement of resources. Specifically, we first propose an input spatial representation distillation (ISRD) mechanism to transfer spatial knowledge from large images to student's input module, which can facilitate stable knowledge transfer between CNN and ViT. Then, a Teacher-Assistant-Student (TAS) framework is further established to disentangle pixel distillation into the model compression stage and input compression stage, which significantly reduces the overall complexity of pixel distillation and the difficulty of distilling intermediate knowledge. Finally, we adapt pixel distillation to object detection via an aligned feature for preservation (AFP) strategy for TAS, which aligns output dimensions of detectors at each stage by manipulating features and anchors of the assistant. Comprehensive experiments on image classification and object detection demonstrate the effectiveness of our method.
Förderung
- National Science and Technology Major Project
- National Natural Science Foundation of China
- Anhui Provincial Key R&D Programmes
- University Synergy Innovation Program of Anhui Province
Zitieren
APA
Guo, G., Zhang, D., Han, L., Liu, N., Cheng, M. M., & Han, J. (2024). Pixel Distillation: Cost-Flexible Distillation Across Image Sizes and Heterogeneous Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12), 9536-9550. https://doi.org/10.1109/tpami.2024.3421277
Vancouver
Guo G, Zhang D, Han L, Liu N, Cheng MM, Han J. Pixel Distillation: Cost-Flexible Distillation Across Image Sizes and Heterogeneous Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2024;46(12):9536-9550. doi:10.1109/tpami.2024.3421277
BibTeX
@article{Guo2024,
author = {Guo, Guangyu and Zhang, Dingwen and Han, Longfei and Liu, Nian and Cheng, Ming-Ming and Han, Junwei},
title = {Pixel Distillation: Cost-Flexible Distillation Across Image Sizes and Heterogeneous Networks},
journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
year = {2024},
volume = {46},
number = {12},
pages = {9536--9550},
doi = {10.1109/tpami.2024.3421277}
}
EUVIMED