Vision Transformers Learn Gestalt-Like Figure-Ground Cues

Researchers have investigated how Vision Transformers (ViTs) learn Gestalt-like figure-ground cues from natural images. By fitting linear probes to intermediate patch representations of 25 ViTs, they found that these models robustly encode surroundedness and convexity. Probes trained on natural images could generalize to artificial stimuli isolating these cues across several models. However, the encoding of symmetry was mixed, appearing for uniformly colored regions but not for textured ones. This study suggests that ViTs can learn figure-ground cues from natural scene statistics, positioning them as a valuable system for studying perceptual organization mechanisms. AI

IMPACT Demonstrates that AI models can learn complex perceptual organization principles from natural data, potentially advancing computer vision.

RANK_REASON Research paper detailing findings on Vision Transformer capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• Matthias Tangemann
• ScienceCast
• Vision Transformers

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →