Motivation
Challenges:
- Retrieve all the images depicting the same landmark regardless of visual similarity.
- Ranking is necessary for the MAP metric.
!
Iterative DELF Query Expansion Motivation
Put images with the same landmark closer to the approximated centers of the landmark clusters iteratively.
!
!
- Involve local descriptors and geometric verification.
- Test images can be used as additional bridging.
Google Landmark Retrieval Challenge:
- Largest public dataset for image retrieval.
- 15K unique landmarks, 1M training images, 1M index images, 100K test images.
- Images have various sizes and high resolution, 329GB in total.
- Single model without ensemble.
Validation:
- No fine-tuning.
Highlights:
- No fine-tuning.
Experiment Results:
| Method | PubMAP |
|---|---|
| DIR | 42.3% |
| DIR + ΔQE | 47.9% |
| DIR + ID-QE | 55.7% |
| DIR + ID-QE + MMST-C^ | 62.7% |
Competition Results:
| Team | PvtMAP |
|---|---|
| 1. CVS5P & Visual Atoms | 62.7% |
| 2. Layer 6 AI | 60.8% |
| 3. SevenSpace | 59.8% |
| 4. Naveer Labs Europe | 58.6% |
| 5. VPP | 58.3% |
- Scalability: fast approximate update for both new images in index and in test.
- Flexibility: limit depth to constrain visual similarity.
- Iterative DELF QE constraints the global feature space.
- Clustering resolves the challenge of visually not similar images through transition of bridging images.
- Modified maximum spanning tree algorithm ranks the candidates across the connected images.
- Noh H et al. Large-Scale Image Retrieval with Attentive Deep Local Features. Proc. ICCV 2017.
- Gordo A et al. Deep Image Retrieval: Learning Global