Research map / Learning theory and algorithms / Kernels and regression theory
When VLAD Met Hilbert
Mehrtash Harandi, Mathieu Salzmann, Fatih Porikli
Check 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 |
|---|---|---|---|
| Use kernel VLAD rather than log-Euclidean VLAD for SPD descriptors | VLAD on SPD data · QMUL, HOCoffee, HOC | accuracy: 87.6% -> 92.7%; 77.2 % -> 84.0 %; 79.7% -> 84.1 % | Mixed |
| Use local-subspace approximation sVLAD instead of global Nyström approximation nVLAD | Kernel VLAD approximations · QMUL, HOCoffee, HOC; Ballet, CIFAR, CMU-PIE; VOC 2007; FMD | accuracy / mAP / CCR: 88.9% -> 92.7%; 83.4% -> 84.0 %; 81.4% -> 84.1 %; 88.9% -> 94.4%; 62.2 % -> 65.2 %; 79.5 % -> 80.1%; 56.2% -> 60.3%; 52. | Improved |
| Use kernel VLAD rather than holistic Grassmann sparse coding | Grassmannian image-set recognition · Ballet, CIFAR, CMU-PIE | accuracy: 79.7 % -> 94.4%; 59.9 % -> 67.9 %; 75.5% -> 86.3 % | Improved |
| Use sVLAD rather than conventional VLAD with Euclidean descriptors | VLAD with SIFT descriptors · VOC 2007 | mAP: 54.7% -> 60.3% | Improved |
| Use sVLAD rather than conventional VLAD with Euclidean descriptors | VLAD with SIFT descriptors · FMD | CCR: 49.4% -> 55.2% | Improved |
| Increase the lE-VLAD dictionary to 256 atoms | lE-VLAD · Ballet | accuracy: 91.1 % -> 91.7% | Improved |
| Use enlarged lE-VLAD rather than sVLAD | Grassmannian VLAD · Ballet | accuracy: 91.7% vs 94.4% | Hurt |
| Add signed square-root normalization to sVLAD (ablation) | sVLAD · QMUL | accuracy: roughly 1% | Improved |
| Use sVLAD rather than kVLAD, enabling square-root normalization | Kernel VLAD on SPD data · QMUL | accuracy: 92.2% -> 92.7% | Improved |
| Use sVLAD rather than kVLAD | Kernel VLAD on SPD data · HOCoffee | accuracy: 84.0 % vs 85.3 % | Hurt |
| Use kVLAD rather than sVLAD | Kernel VLAD on Grassmannian data · CMU-PIE | accuracy: 80.1% -> 86.3 % | Improved |
| Use sVLAD rather than kVLAD | Kernel VLAD on Grassmannian data · Ballet | accuracy: 92.2 % -> 94.4% | Improved |
Closest papers by meaning
- VLAD-BuFF: Burst-aware Fast Feature Aggregation for Visual Place Recognition (ECCV 2024 · similarity 0.88)
- DeepKSPD: Learning Kernel-matrix-based SPD Representation for Fine-grained Image Recognition (ECCV 2018 · similarity 0.88)
- Improving Large-Scale Image Retrieval Through Robust Aggregation of Local Descriptors (TPAMI 2016 · similarity 0.87)
- Approximate Log-Hilbert-Schmidt Distances Between Covariance Operators for Image Classification (CVPR 2016 · similarity 0.86)
- SuperVLAD: Compact and Robust Image Descriptors for Visual Place Recognition (NeurIPS 2024 · similarity 0.85)
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.84)
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.
