Parallel tile inspection and adaptive zoom improve high-resolution visual search

A new framework, VPS, uses question-conditioned tile readers and a learned zoom controller to outperform sequential zoom-only search across five benchmarks and three model sizes.

Chinese Tech
Xijia Tao · Yihua Teng · Xinyu Fu · Cheng Gong · Ziru Liu · Xudong Xie · +2 more

The University of Hong Kong · Huawei Research · City University of Hong Kong · Huazhong University of Science and Technology

Research Digest··3 min read
Tao et al.

The authors propose VPS, which combines two operations: grid_search inspects a regular grid of image tiles in parallel using question-conditioned sub-agents, and zoom_in adaptively selects a precise or merged region for higher-resolution inspection.

Why this paper

From Huawei Research and 3 others

In one line

Parallel tile inspection with adaptive zoom beats sequential zoom for high-resolution visual question answering, improving mean accuracy in 14 of 15 same-model comparisons.

What we could check

  • ·No code link found
  • ·No weights link found
  • ·No dataset link found
  • ·No compute details found
  • ·No stated limitations found
  • ✓Reports numbers on named benchmarks (2 benchmarks)

Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.

§
newspaper

Research Digest

Articles published under the Zotpaper byline are synthesized from multiple source publications by our AI editor and reviewed by our editorial process. Each story combines reporting from credible outlets to give readers a balanced, comprehensive view.