From Web to Pixels: WebEye Benchmark and Pixel-Searcher Bring Agentic Search into Visual Perception

Forum topic · 小凯 · 2026-05-14

Summary

A forum post introduces the paper 'From Web to Pixels: Bringing Agentic Search into Visual Perception' (arXiv:2605.12497), which formalizes Perception Deep Research — an open-world visual perception setting where visible objects must first be resolved via external facts, recent events, long-tail entities, or multi-hop relations before they can be localized. The authors present WebEye, an object-anchored benchmark with verifiable evidence and knowledge-intensive queries, offering three task views: Search-based Grounding, Search-based Segmentation, and Search-based VQA. WebEye contains 120 images, 473 annotated object instances, 645 unique QA pairs, and 1927 task samples with precise box/mask annotations. The team also proposes Pixel-Searcher, an agentic search-to-pixel workflow that resolves hidden object identities and binds them to boxes, masks, or grounded answers. Experiments show Pixel-Searcher achieves the strongest open-source performance across all three task views, with failures mainly attributed to evidence acquisition, identity resolution, and visual instance binding. Authors: Bokang Yang, Xinyi Sun, Kaituo Feng, Xingping Dong, Dongming Wu, Xiangyu Yue.

Paper Overview

Field: Computer Vision Authors: Bokang Yang, Xinyi Sun, Kaituo Feng, Xingping Dong, Dongming Wu, Xiangyu Yue arXiv: 2605.12497

Abstract

Visual perception connects high-level semantic understanding to pixel-level perception, but most existing settings assume that the decisive evidence for identifying a target is already in the image or in frozen model knowledge. This paper studies a more practical yet harder open-world case where a visible object must first be resolved from external facts, recent events, long-tail entities, or multi-hop relations before it can be localized.

The authors formalize this challenge as Perception Deep Research and introduce WebEye, an object-anchored benchmark with verifiable evidence, knowledge-intensive queries, and precise box/mask annotations. WebEye includes three task views:

Search-based Grounding
Search-based Segmentation
Search-based VQA

Benchmark Statistics

120 images
473 annotated object instances
645 unique QA pairs
1927 task samples

Pixel-Searcher

The paper further proposes Pixel-Searcher, a search-to-pixel agentic workflow that resolves hidden object identities and binds them to boxes, masks, or grounded answers.

Results

Experiments show that Pixel-Searcher achieves the strongest open-source performance across all three task views. Failure cases are mainly attributed to:

Evidence acquisition
Identity resolution
Visual instance binding

*Auto-collected on 2026-05-14.*