Research map / Label-efficient and robust learning / Self-supervised and contrastive
Replacing Labeled Real-Image Datasets With Auto-Generated Contours
Hirokatsu Kataoka, Ryo Hayamizu, Ryosuke Yamada, Kodai Nakashima, Sora Takashima, Xinyu Zhang, Edgar Josafat Martinez-Noriega, Nakamasa Inoue, Rio Yokota
Check 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 |
|---|---|---|---|
| pre-train on ExFractalDB-21k instead of ImageNet-21k | ViT-Base · ImageNet-1k | top-1 accuracy: 81.8 -> 82.7 | Improved |
| pre-train on RCDB-21k instead of ImageNet-21k | ViT-Base · ImageNet-1k | top-1 accuracy: 81.8 -> 82.4 | Improved |
| pre-train on ExFractalDB-21k instead of ImageNet-21k | Swin Transformer · COCO detection | AP: 48.8 -> 48.0 | Hurt |
| pre-train on ExFractalDB-21k instead of ImageNet-21k | Swin Transformer · COCO instance segmentation | AP: 43.6 -> 42.8 | Hurt |
| use ExFractalDB-1k pre-training instead of ImageNet-1k supervised pre-training | fine-tuned models · Stanford Cars | accuracy: 89.9 -> 90.3 | Improved |
| use ExFractalDB-1k* pre-training instead of ImageNet-1k supervised pre-training | fine-tuned models · Flowers | accuracy: 99.4 -> 99.6 | Improved |
| increase the number of varied fractal parameters from three to six (ablation) | FractalDB pre-training · CIFAR-10 | accuracy: 96.8 -> 97.0 | Improved |
| use contour vertex counts of 203–502 instead of 3–102 (ablation) | RCDB pre-training · CIFAR-10 | accuracy: 95.5 -> 71.3; 95.5 -> 59.4; 95.5 -> 40.1 | Hurt |
| add 1k background-colored lines to break object contours (ablation) | RCDB pre-training · CIFAR-10 | accuracy: 95.5 -> 46.2 | Hurt |
| use a line-counting task with consistent labels instead of randomly permuted labels (ablation) | LineDB pre-training · CIFAR-10 | accuracy: 13.5 -> 95.6 | Improved |
| increase the number of line-counting categories from 512 to 1k (ablation) | LineDB pre-training · CIFAR-10 | accuracy | Hurt |
Closest papers by meaning
- Visual Atoms: Pre-Training Vision Transformers With Sinusoidal Waves (CVPR 2023 · similarity 0.94)
- Pre-training Vision Transformers with Very Limited Synthesized Images (ICCV 2023 · similarity 0.92)
- Pre-training Vision Models with Mandelbulb Variations (CVPR 2024 · similarity 0.90)
- Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 2024 · similarity 0.85)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.85)
- Should VLMs be Pre-trained with Image Data? (ICLR 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.
