Research map / Architectures and efficiency / Pruning and efficient tuning
Fine-Tuning CNN Image Retrieval with No Human Annotation
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 |
|---|---|---|---|
| replace max pooling (MAC) with learned generalized-mean pooling (GeM) | fine-tuned AlexNet · Oxford5k | mAP: 62.2 -> 67.7 | Improved |
| replace average pooling (SPoC) with learned GeM | fine-tuned AlexNet · Paris6k | mAP: 70.8 -> 75.5 | Improved |
| learn one pooling parameter per feature map instead of a single shared p (ablation) | GeM AlexNet · Oxford5k | mAP: 67.7 -> 66.3 | Hurt |
| learn the shared GeM parameter instead of fixing p=3 (ablation) | GeM AlexNet · Paris6k | mAP: 74.8 -> 75.5 | Improved |
| add learned whitening Lw instead of PCA whitening (ablation) | fine-tuned VGG GeM · Paris6k | mAP: 84.5 -> 86.0 | Improved |
| add PCA whitening to raw descriptors (ablation) | fine-tuned VGG MAC/GeM · Oxford5k (AlexNet MAC) | mAP: 62.2 -> 56.9 | Hurt |
| learn whitening end-to-end inside the network instead of post-hoc Lw (ablation) | AlexNet MAC · Oxford105k | mAP: 52.8 -> 49.6 | Hurt |
| replace average pooling of multi-scale descriptors with GeM (ablation) | VGG GeM, Lw · Oxford5k | mAP: 87.2 -> 87.9 | Improved |
| add multi-scale representation (scales 1, 1/sqrt2, 1/2) at test time (ablation) | VGG GeM, Lw · Oxford5k | mAP: 85.9 -> 87.9 | Improved |
| add alpha-weighted query expansion | VGG GeM multi-scale Lw · Oxford5k | mAP: 87.9 -> 91.9 | Improved |
| train with models containing Oxford and Paris landmarks (ablation) | fine-tuned network · Oxford5k and Paris6k | mAP: +0.3 | No clear effect |
| reduce descriptor dimensionality from 512 to 128 with Lw (ablation) | VGG GeM multi-scale · Oxford5k | mAP: 87.9 -> 81.6 | Hurt |
Closest papers by meaning
- CNN Image Retrieval Learns from BoW: Unsupervised Fine-Tuning with Hard Examples (ECCV 2016 · similarity 0.95)
- End-to-End Learning of Deep Visual Representations for Image Retrieval (IJCV 2017 · similarity 0.90)
- Deep Image Retrieval: Learning Global Representations for Image Search (ECCV 2016 · similarity 0.90)
- Cross-Convolutional-Layer Pooling for Image Recognition (TPAMI 2016 · similarity 0.89)
- Convolutional Patch Representations for Image Retrieval: An Unsupervised Approach (IJCV 2016 · similarity 0.88)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · 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.
