Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Clustering for Unsupervised Learning of Visual Features
Mathilde Caron, Piotr Bojanowski, Armand Joulin, Matthijs Douze
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 ImageNet instead of uncurated YFCC100M images (ablation) | DeepCluster · Pascal VOC 2007 classification (fc6-8 only) | mAP: 67.3 -> 72.0 | Improved |
| pre-train on uncurated YFCC100M (ablation) | DeepCluster · Pascal VOC transfer | mAP: up to +4.3% on classification and +4.5% on semantic segmentation over prior art | No clear effect |
| use Sobel preprocessing instead of raw RGB input (ablation) | DeepCluster · ImageNet linear probe conv5 | accuracy (%): 30.6 -> 38.2 | Improved |
| use k-means instead of PIC graph clustering (ablation) | DeepCluster · Pascal VOC 2007 classification (fc6-8 only) | mAP: 71.0 -> 72.0 | Improved |
| use PIC instead of k-means clustering (ablation) | DeepCluster · Pascal VOC 2007 detection (fc6-8 only) | mAP: 51.4 -> 53.6 | Mixed |
| use k-means instead of PIC on ImageNet MLP retraining (ablation) | DeepCluster · ImageNet classification (fc6-8 retrained) | Acc@1: 44.0 -> 45.9 | Hurt |
| use a deeper VGG-16 instead of AlexNet (ablation) | DeepCluster · Pascal VOC 2007 detection | mAP: 55.4 -> 65.9 | Improved |
| average over 10 crops instead of a single crop | DeepCluster · ImageNet linear probe conv5 | accuracy (%): 36.1 -> 38.2 | Improved |
| apply unsupervised pretraining with DeepCluster instead of random init | DeepCluster · Pascal VOC 2007 classification (fc6-8 only) | mAP: 29.0 -> 70.4 | Improved |
| use Sobel filtering on supervised VGG-16 input | supervised VGG-16 · Oxford5K instance retrieval | mAP: 5.5 points mAP gain on Oxford, not on Paris | Improved |
| train longer than 400 epochs | DeepCluster · Pascal VOC classification (no fine-tuning) | mAP | No clear effect |
Closest papers by meaning
- Joint Unsupervised Learning of Deep Representations and Image Clusters (CVPR 2016 · similarity 0.91)
- Deep Clustering via Joint Convolutional Autoencoder Embedding and Relative Entropy Minimization (ICCV 2017 · similarity 0.90)
- DeepDPM: Deep Clustering With an Unknown Number of Clusters (CVPR 2022 · similarity 0.89)
- Deep Semantic Clustering by Partition Confidence Maximisation (CVPR 2020 · similarity 0.89)
- Mini-cluster Guided Long-tailed Deep Clustering (ICLR 2026 · similarity 0.88)
- Image Clustering via the Principle of Rate Reduction in the Age of Pretrained Models (ICLR 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.
