Research map / Detection and segmentation / Open-set and few-shot detection
Local Aggregation for Unsupervised Learning of Visual Embeddings
Chengxu Zhuang, Alex Lin Zhai, Daniel Yamins
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 |
|---|---|---|---|
| use dynamic nearest-neighbor background N_4096 instead of all inputs (ablation) | LA · ImageNet NN validation (ResNet-18) | NN accuracy: 30.2 -> 35.7 | Improved |
| use dynamic nearest-neighbor background instead of fixed cluster-based background (ablation) | LA · ImageNet NN validation (ResNet-18) | NN accuracy: 33.2 -> 35.7 | Improved |
| use cluster-based close neighbors instead of nearest-neighbor close neighbors (ablation) | LA · ImageNet NN validation (ResNet-18) | NN accuracy: 0.1 -> 35.7 | Improved |
| use clustered close neighbors instead of only the point itself (ablation) | LA · ImageNet NN validation (ResNet-18) | NN accuracy: 33.9 -> 35.7 | Improved |
| increase number of clusterings H from 1 to 3 (ablation) | LA · ImageNet NN validation (ResNet-18) | NN accuracy: 35.7 -> 36.2 | Improved |
| increase number of clusterings H from 3 to 10 (ablation) | LA · ImageNet NN validation (ResNet-18) | NN accuracy: 36.2 -> 36.1 | No clear effect |
| increase number of clusters m from 10k to 30k with H=10 (ablation) | LA · ImageNet NN validation (ResNet-18) | NN accuracy: 36.1 -> 37.9 | Improved |
| replace k-means with DBSCAN for clustering (ablation) | LA · ImageNet | representation quality | Hurt |
| change background neighbor count k from 4096 to 2048 or 8192 (ablation) | LA · ImageNet NN validation (ResNet-18) | NN accuracy: 35.7 -> 35.4 (k=2048), 35.8 (k=8192) | No clear effect |
| update clusterings more often (every 1000 steps) (ablation) | LA · ImageNet NN validation (ResNet-18) | NN accuracy: 35.7 -> 35.7 | No clear effect |
| cluster center-crop network outputs instead of memory bank (ablation) | LA · ImageNet NN validation (ResNet-18) | NN accuracy: 35.7 -> 35.8 | No clear effect |
| replace Instance Recognition objective with Local Aggregation on ResNet-50 | ResNet-50 · ImageNet linear readout conv5 | top-1 accuracy: 54.0 -> 60.2 | Improved |
Closest papers by meaning
- Unsupervised Learning From Video With Deep Neural Embeddings (CVPR 2020 · similarity 0.89)
- Unsupervised Representation Learning by Predicting Image Rotations (ICLR 2018 · similarity 0.89)
- Beyond Supervised vs. Unsupervised: Representative Benchmarking and Analysis of Image Representation Learning (CVPR 2022 · similarity 0.88)
- Unsupervised Learning From Video to Detect Foreground Objects in Single Images (ICCV 2017 · similarity 0.88)
- Unsupervised Learning of Foreground Object Segmentation (IJCV 2019 · similarity 0.88)
- Unsupervised Deep Learning by Neighbourhood Discovery (ICML 2019 · similarity 0.88)
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.
