Research map / Detection and segmentation / Weakly supervised detection and segmentation
Exploring the Limits of Weakly Supervised Pretraining
Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, Laurens van der Maaten
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 |
|---|---|---|---|
| pretrain on Instagram hashtags (1.5k vocab, ~1B images) instead of IN-1k only | ResNeXt-101 32x16d · ImageNet-1k val | top-1 accuracy: an improvement of 4.6%; 84.2% | Improved |
| increase model size from 32x16d to 32x32d and 32x48d (ablation) | IG-940M-1.5k pretrained ResNeXt-101 · ImageNet-1k val | top-1 accuracy: 84.2% -> 85.1% and 85.4% | Improved |
| increase model size beyond 32x16d when training from scratch on IN-1k (ablation) | ResNeXt-101 · ImageNet-1k val | top-1 accuracy: around 79.6% | No clear effect |
| use 17k hashtags instead of 1.5k hashtags (ablation) | IG-1B pretraining · IN-9k | top-1 accuracy: ~7% | Improved |
| use 17k hashtags instead of 1.5k hashtags (ablation) | IG-1B pretraining · IN-1k | top-1 accuracy | Hurt |
| use uniform or square-root sampling instead of natural sampling of hashtags (ablation) | IG-1B-17k ResNeXt-101 32x4d · ImageNet 1k/5k/9k | top-1 accuracy: 5 to 6% | Improved |
| inject 10% random hashtag noise (ablation) | IG-1B-17k ResNeXt-101 32x16d · ImageNet 1k/5k/9k | top-1 accuracy: loss of less than 1% | Hurt |
| inject 25% random hashtag noise (ablation) | IG-1B-17k ResNeXt-101 32x16d · ImageNet 1k/5k/9k | top-1 accuracy: around 2% | Hurt |
| scale pretraining images up (3.5M to 3.5B) (ablation) | ResNeXt-101 linear classifier · ImageNet and CUB2011 | top-1 accuracy: near log-linear gain | Improved |
| use IG-1B-1k pretraining instead of IN-1k pretraining for keypoint detection | Mask R-CNN · COCO keypoint | keypoint AP: 67.0% -> 65.3% | Hurt |
| use 3.5B-image 17k hashtag pretraining instead of IN-1M-1k | ResNeXt-101 32x16d · ImageNet-1k val | rectified top-1 accuracy: 87.5% -> 90.4% | Improved |
| use per-hashtag sigmoid outputs with binary logistic loss instead of softmax (ablation) | hashtag pretraining · ImageNet transfer | accuracy | Hurt |
Closest papers by meaning
- Revisiting Weakly Supervised Pre-Training of Visual Perception Models (CVPR 2022 · similarity 0.94)
- Hard Mixtures of Experts for Large Scale Weakly Supervised Vision (CVPR 2017 · similarity 0.88)
- PreDet: Large-Scale Weakly Supervised Pre-Training for Detection (ICCV 2021 · similarity 0.88)
- ClusterFit: Improving Generalization of Visual Representations (CVPR 2020 · similarity 0.88)
- Revisiting the Transferability of Supervised Pretraining: An MLP Perspective (CVPR 2022 · similarity 0.88)
- Beyond Random Augmentations: Pretraining with Hard Views (ICLR 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.
