Research map / Architectures and efficiency / CNN design
Efficient Compact Bilinear Pooling via Kronecker Product
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 RM projections with two-level Kronecker product factorization | compact bilinear pooling with VGG16 · FGVC-Aircraft, CUB-200-2011, MIT scene dataset, and DTD | accuracy: 91.2 -> 91.4; 85.6 -> 86.0; 80.3 -> 80.5; 67.0 -> 68.2 | Helped |
| replace RM projections with two-level Kronecker product factorization | compact bilinear pooling with ResNet50 · FGVC-Aircraft, CUB-200-2011, MIT scene dataset, and DTD | accuracy: 92.0 -> 92.1; 85.9 -> 85.7; 83.2 -> 84.1; 70.4 -> 71.4 | Mixed |
| replace RM with TKPF | bilinear pooling layer with VGG16 · inference, batch of 64 | bilinear pooling latency: 87ms -> 3ms | Helped |
| replace RM with TKPF | bilinear pooling layer with ResNet50 · inference, batch of 64 | bilinear pooling latency: 161ms -> 3ms | Helped |
| increase a=b from 32 to 64 (ablation) | TKPF with VGG16 · FGVC-Aircraft | accuracy: 90.6 -> 91.5 | Helped |
| increase a=b from 64 to 96 (ablation) | TKPF with VGG16 · CUB-200-2011, MIT scene dataset, and DTD | accuracy: 85.8 -> 86.0; 79.7 -> 80.5; 67.3 -> 68.2 | Helped |
| increase a=b from 96 to 128 (ablation) | TKPF with VGG16 · FGVC-Aircraft, CUB-200-2011, MIT scene dataset, and DTD | accuracy: 91.4 -> 90.8; 86.0 -> 85.8; 80.5 -> 80.0; 68.2 -> 68.0 | Mixed |
| increase identity size r from 4 to 32 (ablation) | TKPF with VGG16 · FGVC-Aircraft, CUB-200-2011, MIT scene dataset, and DTD | accuracy: 90.8 -> 91.4; 85.5 -> 86.0; 79.4 -> 80.5; 67.4 -> 68.2 | Mixed |
| increase the number of Kronecker-product terms from Q=1 to Q=2 (ablation) | TKPF with VGG16 · FGVC-Aircraft, CUB-200-2011, MIT scene dataset, and DTD | accuracy: 90.9 -> 91.4; 85.8 -> 86.0; 79.6 -> 80.5; 67.1 -> 68.2 | Helped |
| increase the number of Kronecker-product terms from Q=2 to Q=4 (ablation) | TKPF with VGG16 · FGVC-Aircraft, CUB-200-2011, MIT scene dataset, and DTD | accuracy: 91.4 -> 91.2; 86.0 -> 85.9; 80.5 -> 79.7; 68.2 -> 68.3 | Mixed |
| use a second-level Kronecker-product summation instead of the single-entry form (ablation) | TKPF · fine-grained recognition benchmarks | accuracy | No clear effect |
| use the symmetric factorization structure instead of the asymmetric structure (ablation) | TKPF · fine-grained recognition benchmarks | accuracy | Hurt |
Closest papers by meaning
- Compact Bilinear Pooling (CVPR 2016 · similarity 0.92)
- Fast and Compact Bilinear Pooling by Shifted Random Maclaurin (AAAI 2021 · similarity 0.90)
- Low-Rank Bilinear Pooling for Fine-Grained Classification (CVPR 2017 · similarity 0.90)
- Revisiting Bilinear Pooling: A Coding Perspective (AAAI 2020 · similarity 0.87)
- SkipPool: Improved Sparse Hierarchical Graph Pooling with Differentiable Exploration (AAAI 2025 · similarity 0.79)
- The Balanced-Pairwise-Affinities Feature Transform (ICML 2024 · similarity 0.79)
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.