AI Paper Atlas

Research map / Label-efficient and robust learning / Semi-supervised learning

Species196: A One-Million Semi-supervised Dataset for Fine-grained Species Recognition

Wei He, Kai Han, Ying Nie, Chengcheng Wang, Yunhe Wang

NeurIPS 2023 · Label-efficient and robust learning › Semi-supervised learning · approach: Holistic semi-supervised pipelines · 2 citations (OpenAlex)

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 changedWhereMeasuredEffect
Train a larger ResNet34 student from a ResNet18 teacher using Noisy StudentNoisy Student on Species196-U · Species196-Lstudent top-1 accuracy: 71.1 -> 73.4Improved
Use a same-size ResNet18 student instead of a larger student (ablation)Noisy Student on Species196-U · Species196-Lstudent top-1 accuracy: 71.1 -> 70.6Hurt
Pretrain ViT-B with SimMIM on Species-U rather than ImageNet-1K (ablation)ViT-B · Species196-Ltop-1 accuracy: 80.9 -> 81.0Improved
Pretrain Swin-B with SimMIM on Species-U rather than ImageNet-1K (ablation)Swin-B · Species196-Ltop-1 accuracy: 80.5 -> 81.6Improved
Pretrain ResNet50 with SparK on Species-U rather than ImageNet-1K (ablation)ResNet50 · Species196-Ltop-1 accuracy: 73.2 -> 73.2No clear effect
Pretrain ConvNeXt V2-T with FCMAE on Species-U rather than ImageNet-1K (ablation)ConvNeXt V2-T · Species196-Ltop-1 accuracy: 77.1 -> 79.3Improved
Pretrain ConvNeXt V2-B with FCMAE on Species-U rather than ImageNet-1K (ablation)ConvNeXt V2-B · Species196-Ltop-1 accuracy: 79.9 -> 81.2Improved
Use the Species-U domain-specific data for masked-image pretraining instead of ImageNet-1K (ablation)MIM-pretrained models · Species196-Ltop-1 accuracy: ViT-B: 80.9 -> 81.0; Swin-B: 80.5 -> 81.6; ResNet50: 73.2 -> 73.2; ConvNeXt V2-T: 77.1 -> 79.3; ConvNeXt V2-B: 79.9 -> 8Mixed
Use a larger Noisy Student model rather than a same-size student (ablation)Noisy Student · Species196-Lstudent top-1 accuracy: ResNet18 teacher to ResNet34 student: 71.1 -> 73.4; ResNet18 teacher to ResNet18 student: 71.1 -> 70.6; ResNet34 teacherMixed
Combine class-name and species-description prompts (1) + (4) + (5) (ablation)OpenCLIP ViT-B/32 · Species196-L zero-shot classificationtop-1 accuracy: 10.7 -> 10.94Improved
Add prompt (6) to combination (1) + (4) + (5) (ablation)OpenCLIP ViT-B/32 · Species196-L zero-shot classificationtop-1 accuracy: 10.94 -> 10.91Hurt
Use common names rather than scientific names as classnames (ablation)OpenCLIP models · Species196-L zero-shot classificationtop-1 accuracy: ViT-B/32: 10.94 -> 16.71; ViT-B/16: 10.40 -> 16.90; ViT-L/14: 12.07 -> 18.43; ViT-H/14: 15.00 -> 22.57Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy