Research map / Label-efficient and robust learning / Recognition and long-tail
Visual Recognition with Deep Nearest Centroids
Wenguan Wang, Cheng Han, Tianfei Zhou, Dongfang Liu
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 |
|---|---|---|---|
| Replace a parametric softmax classifier with DNC | ResNet classification models · CIFAR-10 test | top-1 accuracy: 95.55% -> 95.78%; 95.58% -> 95.82% | Improved |
| Replace a parametric softmax classifier with DNC | ResNet and Swin classification models · ImageNet val | top-1 accuracy: 76.20% -> 76.49%; 77.52% -> 77.80%; 83.02% -> 83.26%; 83.36% -> 83.68% | Improved |
| Replace a parametric softmax classifier with DNC | ResNet classification models · CIFAR-100 test | top-1 accuracy: 79.81% -> 79.91%; 79.83% -> 79.99% | Improved |
| Use DNC classification and an ImageNet-pretrained DNC backbone | FCN with ResNet101 · ADE20K val | mIoU: 39.9% -> 42.3% | Improved |
| Use DNC classification and an ImageNet-pretrained DNC backbone | FCN with ResNet101 · Cityscapes val | mIoU: 75.6% -> 77.5% | Improved |
| Use DNC classification and an ImageNet-pretrained DNC backbone | DeepLabV3 with ResNet101 · ADE20K val | mIoU: 44.1% -> 45.7% | Improved |
| Use DNC classification and an ImageNet-pretrained DNC backbone | UperNet with Swin-B · ADE20K val | mIoU: 48.0% -> 50.5% | Improved |
| Anchor class sub-centroids to actual training images instead of cluster centers (ablation) | DNC-ResNet50 · ImageNet val | top-1 accuracy: 76.49% -> 76.37% | Hurt |
| Increase class representatives from one centroid to four sub-centroids (ablation) | DNC-ResNet101 · ImageNet val | top-1 accuracy: 77.31% -> 77.80% | Improved |
| Increase class representatives from one centroid to ten sub-centroids (ablation) | DNC-DeepLabV3 · ADE20K val | mIoU: 43.2% -> 44.3% | Improved |
| Increase the number of class sub-centroids from ten to twenty (ablation) | DNC-DeepLabV3 · ADE20K val | mIoU: 44.3% -> 44.0% | Hurt |
| Increase external memory size (ablation) | DNC-ResNet101 · ImageNet val | top-1 accuracy: 77.49% -> 77.80% | Improved |
Closest papers by meaning
- ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models (NeurIPS 2019 · similarity 0.89)
- Bilinear Convolutional Neural Networks for Fine-Grained Visual Recognition (TPAMI 2017 · similarity 0.88)
- Improving Generalization via Scalable Neighborhood Component Analysis (ECCV 2018 · similarity 0.88)
- Deep Residual Learning for Image Recognition (CVPR 2016 · similarity 0.87)
- Large Scale Visual Recognition through Adaptation using Joint Representation and Multiple Instance Learning (JMLR 2016 · similarity 0.87)
- Is Second-Order Information Helpful for Large-Scale Visual Recognition? (ICCV 2017 · 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.
