EUVIMEDEUVIMED entdecken
Crossref

P‐52: Distinguished Student & Invited Poster Paper: XStreamVGGT: Extremely Memory‐Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compression

Zunhai Su, Weihao Ye, Hansen Feng, Keyu Fan, Jing Zhang, Dahai Yu, Zhengwu Liu, Ngai Wong

SID Symposium Digest of Technical Papers · 2026-06 · 57 (1) · 1865-1868

DOI: 10.1002/sdtp.70001_506 ↗

Kurz gefasst

Fachartikel in der Zeitschrift SID Symposium Digest of Technical Papers, online erschienen am 11.09.2026. Verfasst von Zunhai Su und 7 weiteren Autor:innen. Die Arbeit nennt 20 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Learning‐based 3D reconstruction has advanced through transformers like StreamVGGT, which enables streaming reconstruction via frame‐wise causal attention but suffers unbounded KV cache growth. We propose XStreamVGGT, a tuning‐free approach compressing KV cache through joint pruning and quantization. By pruning redundant tokens and applying distribution‐aware quantization, XStreamVGGT achieves 4.42× memory reduction and 5.48× speedup with mostly negligible performance degradation, enabling scalable streaming applications.

Zitieren

APA

Su, Z., Ye, W., Feng, H., Fan, K., Zhang, J., Yu, D., Liu, Z., & Wong, N. (2026). P‐52: Distinguished Student & Invited Poster Paper: XStreamVGGT: Extremely Memory‐Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compression. SID Symposium Digest of Technical Papers, 57(1), 1865-1868. https://doi.org/10.1002/sdtp.70001_506

Vancouver

Su Z, Ye W, Feng H, Fan K, Zhang J, Yu D, et al. P‐52: Distinguished Student & Invited Poster Paper: XStreamVGGT: Extremely Memory‐Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compression. SID Symposium Digest of Technical Papers. 2026;57(1):1865-1868. doi:10.1002/sdtp.70001_506

BibTeX
@article{Su2026,
  author = {Su, Zunhai and Ye, Weihao and Feng, Hansen and Fan, Keyu and Zhang, Jing and Yu, Dahai and Liu, Zhengwu and Wong, Ngai},
  title = {P‐52: Distinguished Student & Invited Poster Paper: XStreamVGGT: Extremely Memory‐Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compression},
  journal = {SID Symposium Digest of Technical Papers},
  year = {2026},
  volume = {57},
  number = {1},
  pages = {1865--1868},
  doi = {10.1002/sdtp.70001_506}
}
Kostenlos starten →