Research map / Image generation / Autoregressive and latent generation
SketchEmbedNet: Learning Novel Concepts by Imitating Drawings
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| add pixel loss with curriculum weighting (alpha_max=0.50 vs 0) (ablation) | SketchEmbedNet on Omniglot · Omniglot (20,1) | accuracy: +4.5 points (87.17 to 91.67) | Helped |
| raise alpha_max to 1.0 (pixel loss dominates, teacher forcing) (ablation) | SketchEmbedNet pixel loss weighting · Omniglot (20,1) and generated sketch classification | accuracy: Omniglot 87.63 vs 91.67 at 0.50; generated-sketch seen accuracy 4.80 vs 81.44 | Hurt |
| train on Quickdraw sketches instead of Omniglot sketches | SketchEmbedNet · Omniglot (20,1) | accuracy: 91.67 vs 86.18 | Helped |
| use the sketch-generation objective (SketchEmbedNet) instead of contrastive matching of sketch and image embeddings | Quickdraw-trained embeddings · Omniglot (20,1) | accuracy: 91.67 vs 73.01 | Helped |
| use sketch generation objective instead of contrastive baseline on Sketchy-trained model | ResNet12 embeddings trained on Sketchy · mini-ImageNet (5,1) | accuracy: 40.39 vs 30.56 | Helped |
| use sketch-generation objective instead of contrastive matching for novel-class Quickdraw classification | SketchEmbedNet · Quickdraw 45-way unseen classes | accuracy: 75.68 vs 70.72 | Helped |
| use sketch-generation objective (with pixel loss) instead of Conv-VAE on Quickdraw | Quickdraw embeddings · Quickdraw 45-way unseen classes | accuracy: 75.68 vs 53.06 | Helped |
| deeper ResNet12 encoder instead of Conv4 on Sketchy training | SketchEmbedNet trained on Sketchy · mini-ImageNet (5,1) | accuracy: 40.39 vs 38.61 | Helped |
| train with 15 random seeds to measure variance (ablation) | SketchEmbedNet Omniglot (Conv4) (5,1) · Omniglot (5,1) | accuracy: seed average 96.37 ± 0.12 vs best 96.96 | No clear effect |
| add class-label prediction head while sketching (w/ Labels) (ablation) | SketchEmbedNet on Omniglot · Omniglot (20,1) | accuracy: 71.35 vs 90.69 seed average | Hurt |
| use Conv-VAE pixel-based image generation instead of sketch generation | Generated sketch classification · Seen/Unseen classes accuracy | accuracy: 76.28 vs 81.44 (seen); 75.07 vs 77.94 (unseen) | Hurt |
Closest papers by meaning
- A Neural Representation of Sketch Drawings (ICLR 2018 · similarity 0.90)
- SketchNet: Sketch Classification With Web Images (CVPR 2016 · similarity 0.89)
- Doodle It Yourself: Class Incremental Learning by Drawing a Few Sketches (CVPR 2022 · similarity 0.89)
- SketchINR: A First Look into Sketches as Implicit Neural Representations (CVPR 2024 · similarity 0.88)
- It's All About Your Sketch: Democratising Sketch Control in Diffusion Models (CVPR 2024 · similarity 0.87)
- Sketch Down the FLOPs: Towards Efficient Networks for Human Sketch (CVPR 2025 · similarity 0.86)
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.