Research map / Label-efficient and robust learning / Self-supervised and contrastive
Human-like Object Grouping in Self-supervised Vision Transformers
Hossein Adeli, Seoyoung Ahn, Andrew Luo, Mengmi Zhang, Nikolaus Kriegeskorte, Gregory Zelinsky
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DINO-family self-supervised training instead of supervised training (same ViT-B) | ViT B · Two-dot grouping | grouping accuracy (%): 70.6 -> 91.9 | Improved |
| add large-scale pretraining data (IN21k vs IN1K) | supervised ConvNext B · Two-dot grouping | grouping accuracy (%): 60.0 -> 67.4 | Improved |
| add large-scale pretraining data (IN21k vs IN1K) | supervised ViT B · Human RT prediction | noise-normalized Spearman r: 0.59 -> 0.59 | No clear effect |
| distill DINOv3 ViT into a ConvNext | ConvNext B · Two-dot grouping | grouping accuracy (%): 67.4 -> 86.7 | Improved |
| add Gram matrix distillation from DINOv3 ViT B (ablation) | IN21k ViT B · Human RT prediction | noise-normalized Spearman r: 0.59 -> 0.72 | Improved |
| add Gram matrix distillation from DINOv3 ViT B (ablation) | IN1K ViT B · Human RT prediction | noise-normalized Spearman r: 0.59 -> 0.70 | Improved |
| add Gram matrix distillation from DINOv3 ViT B (ablation) | supervised ViT/ConvNext B (4 models) · Two-dot grouping | grouping accuracy (%): +8 to +18 points | Improved |
| add Gram matrix distillation from DINOv3 ViT B (ablation) | supervised ViT/ConvNext B (4 models) · Object-centric AUC | AUC | Improved |
| replace Gram distillation with CLS token distillation (ablation) | IN21k ViT B · Human RT prediction | noise-normalized Spearman r: 0.59 -> 0.59 | No clear effect |
| replace Gram distillation with CLS token distillation (ablation) | IN1K ViT B · Human RT prediction | noise-normalized Spearman r: 0.59 -> 0.58 | No clear effect |
| add Gram distillation (Transformer vs ConvNext) (ablation) | supervised models · Behavioral alignment | noise-normalized Spearman r: Transformers larger gains than ConvNext | Mixed |
| use higher object-centric AUC across models | 9 vision models · Behavioral alignment | Spearman r: r=0.950 | Improved |
Closest papers by meaning
- Probing the Mid-level Vision Capabilities of Self-Supervised Learning (CVPR 2025 · similarity 0.88)
- Evaluating Multiview Object Consistency in Humans and Image Models (NeurIPS 2024 · similarity 0.88)
- Not Too Generative, Not Too Discriminative: The Human Alignment Sweet Spot (NeurIPS 2026 · similarity 0.88)
- Contour Integration Underlies Human-Like Vision (ICML 2025 · similarity 0.88)
- When does perceptual alignment benefit vision representations? (NeurIPS 2024 · similarity 0.87)
- Vision Transformers Learn Gestalt-Like Figure-Ground Cues from Natural Images (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
