Research map / Label-efficient and robust learning / Recognition and long-tail
Convolutional Patch Representations for Image Retrieval: An Unsupervised Approach
Mattis Paulin, Julien Mairal, Matthijs Douze, Zaïd Harchaoui, Florent Perronnin, Cordelia Schmid
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 |
|---|---|---|---|
| retrain PhilippNet on 100K real Rome classes instead of surrogate classes (ablation) | PhilippNet · RomePatches-train patch retrieval | mAP: 86.1 -> 89.9 | Improved |
| retrain PhilippNet on only 10K Rome classes (ablation) | PhilippNet · RomePatches-train patch retrieval | mAP: 86.1 -> 84.1 | Hurt |
| use CKN-grad unsupervised instead of supervised CNNs | PhilippNet-Rome 100K · RomePatches-test patch retrieval | mAP: 85.6 -> 88.1 | Improved |
| fine-tune AlexNet on Landmarks instead of ImageNet (conv2) (ablation) | AlexNet · Oxford image retrieval | mAP: 12.5 -> 23.8 | Improved |
| fine-tune AlexNet on Landmarks instead of ImageNet (conv3) (ablation) | AlexNet · Holidays image retrieval | mAP: 79.3 -> 76.1 | Hurt |
| fine-tune AlexNet on Landmarks instead of ImageNet (conv3) (ablation) | AlexNet · Oxford image retrieval | mAP: 33.3 -> 38.6 | Improved |
| concatenate VLADs of the three CKN channels (mix) versus CKN-white (ablation) | CKN · Holidays image retrieval | mAP: 78.7 -> 79.3 | Improved |
| PCA+whitening to 4096 dims | CKN-mix · Holidays image retrieval | mAP: 79.3 -> 82.9 | Improved |
| PCA+whitening to 4096 dims | CKN-grad · Oxford image retrieval | mAP: 49.8 -> 47.2 | Hurt |
| use dense keypoints with increased pooling instead of Hessian-Affine | CKN-white · Holidays image retrieval | mAP: 78.7 -> 80.8 | Improved |
| use dense keypoints with unchanged pooling instead of Hessian-Affine (ablation) | CKN-white · Holidays image retrieval | mAP: 78.7 -> 76.8 | Hurt |
| retune CKN-grad hyperparameters (fewer bins, more pooling) for DoG patches | CKN-grad · Liberty pairs | FPR@95% recall: 27.7 -> 14.4 | Improved |
Closest papers by meaning
- Efficient Learning of CNNs using Patch Based Features (ICML 2022 · similarity 0.90)
- CNN Image Retrieval Learns from BoW: Unsupervised Fine-Tuning with Hard Examples (ECCV 2016 · similarity 0.89)
- The Unreasonable Effectiveness of Patches in Deep Convolutional Kernels Methods (ICLR 2021 · similarity 0.89)
- End-to-End Learning of Deep Visual Representations for Image Retrieval (IJCV 2017 · similarity 0.89)
- Patch Reordering: A NovelWay to Achieve Rotation and Translation Invariance in Convolutional Neural Networks (AAAI 2017 · similarity 0.89)
- SIFT Meets CNN: A Decade Survey of Instance Retrieval (TPAMI 2017 · similarity 0.89)
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.
