Research map / Image generation / Autoregressive and latent generation
Efficient Diffusion on Region Manifolds: Recovering Small Objects With Compact CNN Representations
Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, Teddy Furon, Ondrej Chum
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 |
|---|---|---|---|
| use diffusion over region descriptors instead of global image descriptors | CNN image retrieval · INSTRE | mAP: 47.7 -> 80.0 | Improved |
| use diffusion over region descriptors instead of global image descriptors | CNN image retrieval · Oxford5k | mAP: 77.7 -> 93.2 | Improved |
| use diffusion over region descriptors instead of global image descriptors | CNN image retrieval · Oxford105k | mAP: 70.1 -> 90.3 | Improved |
| use diffusion over region descriptors instead of global image descriptors | CNN image retrieval · Paris6k | mAP: 84.1 -> 96.5 | Improved |
| use diffusion over region descriptors instead of global image descriptors | CNN image retrieval · Paris106k | mAP: 76.8 -> 92.6 | Improved |
| use generalized max pooling instead of sum pooling after regional diffusion (ablation) | Regional diffusion · INSTRE; Oxford5k; Oxford105k; Paris6k; Paris106k | mAP: 79.1 -> 80.0; 92.2 -> 93.2; 90.6 -> 91.6; 96.1 -> 96.5; 94.4 -> 94.6 | Improved |
| handle unseen queries by weighting their nearest neighbors without modifying the affinity matrix (ablation) | Regional diffusion · INSTRE | mAP: 77.7 -> 80.0 | Improved |
| reduce each image's regional descriptors with per-image GMMs (ablation) | Regional diffusion · Oxford5k | mAP: 93.2 -> 91.5 | Hurt |
| reduce each image's regional descriptors with per-image GMMs (ablation) | Regional diffusion · Oxford105k | mAP: 90.3 -> 84.7 | Hurt |
| reduce each image's regional descriptors with per-image GMMs (ablation) | Regional diffusion · INSTRE | mAP: 80.0 -> 77.5 | Hurt |
| replace exhaustive affinity-graph construction with Dong et al.'s approximate k-NN graph (ablation) | Regional diffusion · Oxford105k | mAP: 92.5 -> 91.6 | Hurt |
| replace exhaustive affinity-graph construction with Dong et al.'s approximate k-NN graph (ablation) | Regional diffusion · Paris106k | mAP: 95.2 -> 94.6 | Hurt |
Closest papers by meaning
- Regularized Diffusion Process on Bidirectional Context for Object Retrieval (TPAMI 2018 · similarity 0.88)
- Deep Image Retrieval: Learning Global Representations for Image Search (ECCV 2016 · similarity 0.88)
- CNN Image Retrieval Learns from BoW: Unsupervised Fine-Tuning with Hard Examples (ECCV 2016 · similarity 0.88)
- End-to-End Learning of Deep Visual Representations for Image Retrieval (IJCV 2017 · similarity 0.88)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.85)
- Where's Waldo: Diffusion Features For Personalized Segmentation and Retrieval (NeurIPS 2024 · similarity 0.85)
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.
