Research map / Label-efficient and robust learning / Recognition and long-tail
Unifying Deep Local and Global Features for Image Search
Bingyi Cao, André Araujo, Jack Sim
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 |
|---|---|---|---|
| add stop-gradient from attention and reconstruction losses to the backbone (ablation) | DELG joint model (R50) · ROxf Hard | mAP: 42.4 -> 45.1 | Improved |
| add stop-gradient from attention and reconstruction losses to the backbone (ablation) | DELG joint model (R50) · RPar Medium | mAP: 78.9 -> 81.6 | Improved |
| replace softmax loss with ArcFace loss (ablation) | SPoC global feature · ROxf Medium | mAP: 51.2 -> 59.8 | Improved |
| replace SPoC pooling with GeM pooling (ablation) | ArcFace global feature · ROxf Medium | mAP: 59.8 -> 69.3 | Improved |
| train global and local heads jointly without gradient control (ablation) | GeM+ArcFace global feature · RPar Medium | mAP: 82.2 -> 78.9 | Hurt |
| add the local-feature head with stop-gradient (jointly trained) vs separate global model (ablation) | GeM+ArcFace global feature · ROxf Medium | mAP: 69.3 -> 69.7 | No clear effect |
| add stop-gradient (ablation) | naive joint training · activation sparsity of conv4 and conv5 | sparsity: 80% and 97% -> 45% and 88% | Improved |
| increase autoencoder reconstruction weight lambda from 0 to 10, then 20 (ablation) | autoencoder dimensionality reduction · GLD image-pair matching | AP | Mixed |
| add local feature re-ranking with spatial verification | R101-DELG · ROxf+1M Hard | mAP: 30.3 -> 39.3 | Improved |
| add local feature re-ranking | R101-DELG · GLDv2-recognition | microAP@1: 32.0 -> 61.2 | Improved |
| binarize local features (DELG star) | R101-DELG · ROxf+1M Hard | mAP: 39.3 -> 38.5 | Hurt |
| use 7 local scales instead of 3 | R101-DELG re-ranking · ROxf Medium | mAP: 77.5 -> 78.5 | Improved |
Closest papers by meaning
- Large-Scale Image Retrieval With Attentive Deep Local Features (ICCV 2017 · similarity 0.91)
- Deep Image Retrieval: Learning Global Representations for Image Search (ECCV 2016 · similarity 0.91)
- Learning Super-Features for Image Retrieval (ICLR 2022 · similarity 0.90)
- DOLG: Single-Stage Image Retrieval With Deep Orthogonal Fusion of Local and Global Features (ICCV 2021 · similarity 0.90)
- Local Features and Visual Words Emerge in Activations (CVPR 2019 · similarity 0.89)
- End-to-End Learning of Deep Visual Representations for Image Retrieval (IJCV 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.
