Research map / Label-efficient and robust learning / Recognition and long-tail
The Unreasonable Effectiveness of Noisy Data for Fine-Grained Recognition
Jonathan Krause, Benjamin Sapp, Andrew Howard, Howard Zhou, Alexander Toshev, Tom Duerig, James Philbin, Li Fei-Fei
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 |
|---|---|---|---|
| train on raw web search images instead of the CUB ground truth training set | Inception-v3 · CUB-200-2011 | top-1 accuracy (%): 84.4 -> 87.7 | Improved |
| add cross-category filtering of images appearing in multiple categories (CUB) (ablation) | Web training · CUB-200-2011 | top-1 accuracy (%): 87.7 -> 89.0 | Improved |
| add filtering (Birdsnap) (ablation) | Web training · Birdsnap | top-1 accuracy (%): 76.1 -> 78.2 | Improved |
| add filtering (Stanford Dogs) (ablation) | Web training · Stanford Dogs | top-1 accuracy (%): 78.5 -> 78.4 | No clear effect |
| pre-train on the larger L-Bird (10,982 species) before fine-tuning on CUB categories (ablation) | Web (filtered) · CUB-200-2011 | top-1 accuracy (%): 89.0 -> 91.9 | Improved |
| pre-train on L-Bird before fine-tuning on Birdsnap categories (ablation) | Web (filtered) · Birdsnap | top-1 accuracy (%): 78.2 -> 82.8 | Improved |
| pre-train on larger L-Aircraft (409 types) (ablation) | Web (filtered) · FGVC-Aircraft | top-1 accuracy (%): 91.1 -> 90.9 | No clear effect |
| fine-tune additionally on ground truth FGVC training set (ablation) | L-Aircraft · FGVC-Aircraft | top-1 accuracy (%): 90.9 -> 94.5 | Improved |
| fine-tune additionally on Stanford Dogs ground truth training set (ablation) | L-Dog · Stanford Dogs | top-1 accuracy (%): 78.4 -> 84.0 | Improved |
| use 144 test-time crops instead of single crop (CUB, L-Bird) (ablation) | L-Bird · CUB-200-2011 | top-1 accuracy (%): 91.9 -> 92.3 | Improved |
| add second round of active learning (from scratch) (ablation) | Active learning · Stanford Dogs (active learning comparison) | top-1 accuracy (%): 65.8 -> 74.0 | Improved |
| subsample active learning data to size of filtered web data (ablation) | A.L., one round (ft) · Stanford Dogs (active learning comparison) | top-1 accuracy (%): 78.8 vs 78.4 | No clear effect |
Closest papers by meaning
- Webly Supervised Fine-Grained Recognition: Benchmark Datasets and an Approach (ICCV 2021 · similarity 0.93)
- Fine-Grained Recognition: Accounting for Subtle Differences between Similar Classes (AAAI 2020 · similarity 0.91)
- Fine-Grained Recognition in the Wild: A Multi-Task Domain Adaptation Approach (ICCV 2017 · similarity 0.91)
- Fine-grained Classes and How to Find Them (ICML 2024 · similarity 0.89)
- Democratizing Fine-grained Visual Recognition with Large Language Models (ICLR 2024 · similarity 0.88)
- From Coarse to Fine-Grained Open-Set Recognition (CVPR 2024 · similarity 0.87)
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.
