Research map / Vision-language and multimodal / Human and embodied modeling
Learning to See Through a Baby's Eyes: Early Visual Diets Enable Robust Visual Intelligence in Humans and Machines
Yusen Cai, Qing Lin, Bhargava Satya Nunna, Mengmi Zhang
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 the grayscale-to-color staged order instead of shuffled order (ablation) | CDiet (SimCLR-ResNet) · CO3D-C (CO3D-10, SimCLR-ResNet pretrained on CO3D) | mCE: 87.3 -> 84.8 | Improved |
| use the grayscale-to-color staged order instead of reversed order (ablation) | CDiet (SimCLR-ResNet) · CO3D-C (CO3D-10, SimCLR-ResNet pretrained on CO3D) | mCE: 94.1 -> 84.8 | Improved |
| add early color-degraded stages instead of full-color only (ablation) | CDiet (SimCLR-ResNet) · CO3D-C (CO3D-10, SimCLR-ResNet pretrained on CO3D) | mCE: 90.8 -> 84.8 | Improved |
| integrate color, acuity and temporality diets (ablation) | CATDiet (SimCLR-ResNet) · CO3D-C (CO3D-10, SimCLR-ResNet pretrained on CO3D) | mCE: 84.8 -> 72.0 | Improved |
| add temporal continuity objective over non-smooth baseline (ablation) | TDiet (SimCLR-ResNet) · CO3D-C (CO3D-10, SimCLR-ResNet pretrained on CO3D) | mCE | Improved |
| replace temporally adjacent real frames with warped-only positives (Throughout) (ablation) | CombDiet (SimCLR-ResNet) · CO3D-10 | accuracy: 83.0 -> 70.6 | Hurt |
| replace some real adjacent frames with warped views (Adjacent) (ablation) | CombDiet (SimCLR-ResNet) · CO3D-10 | accuracy: 83.0 -> 82.4 | Hurt |
| halve or double temporal stride (ablation) | CombDiet (SimCLR-ResNet) · CO3D-10 | accuracy: 83.0 -> 83.3 | No clear effect |
| lengthen the CATDiet warm-up from 30% to 70% (ablation) | CombDiet (SimCLR-ViT) · CO3D-10 | accuracy: 74.4 -> 73.1 | Hurt |
| use CombDiet instead of standard SSL | SimCLR-ViT · SAYCam test | accuracy: 48.8 -> 54.5 | Improved |
| use CombDiet instead of standard SSL | DINO-ViT · 3D-PC depth order | dAcc: 74.5 -> 78.9 | Improved |
| use ordered CombDiet instead of shuffled order (ablation) | DINO-ResNet · SAYCam test | accuracy: 43.5 -> 59.9 | Improved |
Closest papers by meaning
- BabyVLM-V2: Toward Developmentally Grounded Pretraining and Benchmarking of Vision Foundation Models (CVPR 2026 · similarity 0.88)
- Learning from an Infant’s Visual Experience (AAAI 2024 · similarity 0.87)
- Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning (CVPR 2025 · similarity 0.87)
- BabyVision: Visual Reasoning Beyond Language (ICML 2026 · similarity 0.86)
- Object Concepts Emerge from Motion (NeurIPS 2025 · similarity 0.85)
- Probing the Mid-level Vision Capabilities of Self-Supervised Learning (CVPR 2025 · similarity 0.85)
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.
