Research map / Detection and segmentation / Weakly supervised detection and segmentation
Learning Visual Features from Large Weakly Supervised Data
Armand Joulin, Laurens van der Maaten, Allan Jabri, Nicolas Vasilache
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 end-to-end on Flickr instead of logistic regression on ImageNet-pretrained features | AlexNet word predictor · Flickr 100M word prediction (K=1,000) | precision@10: 8.27 -> 17.98 | Improved |
| train end-to-end on Flickr instead of ImageNet-pretrained features | AlexNet word predictor · Flickr 100M word prediction (K=100,000) | precision@10: 1.61 -> 2.56 | Improved |
| use multiclass logistic loss instead of one-versus-all logistic loss (ablation) | AlexNet, K=1,000 · Flickr 100M word prediction | precision@10: 16.43 -> 17.98 | Improved |
| increase Flickr training set beyond 1 million images (ablation) | AlexNet, K=1,000 · Flickr word prediction and transfer | precision@10: levels out after ~50 million images | Improved |
| switch from AlexNet to GoogLeNet | Flickr-trained features · ImageNet (transfer) | accuracy: 34.93 -> 33.62 | Hurt |
| use Flickr features instead of ImageNet features | AlexNet · Pascal VOC 2007 (transfer) | mAP: 59.8 -> 62.4 | Improved |
| use Flickr features instead of ImageNet features | GoogLeNet · Pascal VOC 2007 (transfer) | mAP: 77.1 -> 73.2 | Hurt |
| use Flickr features instead of ImageNet features | AlexNet · Oxford Flowers (transfer) | accuracy: 80.28 -> 69.72 | Hurt |
| use Flickr features instead of ImageNet features | AlexNet · MIT SUN (transfer) | accuracy: 41.40 -> 42.67 | Improved |
| use Flickr features instead of ImageNet features | AlexNet · MIT Indoor (transfer) | accuracy: 53.82 -> 53.19 | No clear effect |
| concatenate Flickr and ImageNet features | AlexNet · MIT Indoor (transfer) | accuracy: 53.82 -> 58.76 | Improved |
| concatenate Flickr and ImageNet features | GoogLeNet · Pascal VOC 2007 (transfer) | mAP: 77.1 -> 79.0 | Improved |
Closest papers by meaning
- Revisiting Weakly Supervised Pre-Training of Visual Perception Models (CVPR 2022 · similarity 0.88)
- VirTex: Learning Visual Representations From Textual Annotations (CVPR 2021 · similarity 0.87)
- Large Scale Visual Recognition through Adaptation using Joint Representation and Multiple Instance Learning (JMLR 2016 · similarity 0.87)
- Exploring the Limits of Weakly Supervised Pretraining (ECCV 2018 · similarity 0.87)
- Hard Mixtures of Experts for Large Scale Weakly Supervised Vision (CVPR 2017 · similarity 0.87)
- Low-Shot Learning With Large-Scale Diffusion (CVPR 2018 · 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.
