Research map / Architectures and efficiency / Pruning and efficient tuning
CNN Image Retrieval Learns from BoW: Unsupervised Fine-Tuning with Hard Examples
Filip Radenović, Giorgos Tolias, Ondřej Chum
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 |
|---|---|---|---|
| fine-tune the CNN on SfM-derived pairs instead of off-the-shelf | VGG MAC 512D · Oxford5k | mAP: 56.4 -> 79.7 | Improved |
| add learned whitening Lw instead of no post-processing (ablation) | fine-tuned VGG MAC · Oxford5k | mAP: 78.7 -> 79.7 | Improved |
| use learned whitening Lw instead of PCA whitening (ablation) | fine-tuned VGG MAC · Oxford5k | mAP: 76.1 -> 79.7 | Improved |
| add PCA whitening instead of no post-processing (ablation) | fine-tuned AlexNet MAC · Oxford105k | mAP: 54.2 -> 44.1 | Hurt |
| add PCA whitening instead of no post-processing (ablation) | fine-tuned VGG MAC · Oxford5k / Paris6k | mAP: 78.7 -> 76.1 (Oxford5k); 77.1 -> 79.0 (Paris6k) | Mixed |
| select positives by relaxed inliers (m3) instead of MAC distance (m1) (ablation) | siamese MAC · Oxford105k / Paris106k | mAP | Improved |
| select negatives with at most one per cluster (N2) instead of kNN (N1) (ablation) | siamese MAC · Oxford105k / Paris106k | mAP | Improved |
| use contrastive loss instead of triplet loss (ablation) | siamese MAC · Oxford105k / Paris106k | mAP | Improved |
| train with all 551 3D models instead of 10 or 100 (ablation) | AlexNet MAC · Oxford105k / Paris106k | mAP | Improved |
| include Oxford and Paris landmarks in the training reconstructions (ablation) | fine-tuned MAC · Oxford/Paris | mAP: +0.3 | No clear effect |
| replace MAC with R-MAC regional aggregation | fine-tuned VGG 512D · Oxford5k / Paris6k | mAP: 79.7 -> 77.0 (Oxford5k); 82.4 -> 83.8 (Paris6k) | Mixed |
| add re-ranking and query expansion to MAC | fine-tuned VGG MAC · Oxford5k | mAP: 79.7 -> 85.0 | Improved |
Closest papers by meaning
- Fine-Tuning CNN Image Retrieval with No Human Annotation (TPAMI 2018 · similarity 0.95)
- End-to-End Learning of Deep Visual Representations for Image Retrieval (IJCV 2017 · similarity 0.91)
- Deep Image Retrieval: Learning Global Representations for Image Search (ECCV 2016 · similarity 0.90)
- SIFT Meets CNN: A Decade Survey of Instance Retrieval (TPAMI 2017 · similarity 0.90)
- Convolutional Patch Representations for Image Retrieval: An Unsupervised Approach (IJCV 2016 · similarity 0.89)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · 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.
