Research map / Detection and segmentation / CNN features
Bilinear Convolutional Neural Networks for Fine-Grained Visual Recognition
Tsung‐Yu Lin, Aruni RoyChowdhury, Subhransu Maji
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 domain-specific fine-tuning to NetVLAD | NetVLAD (VGG-D) · CUB-200-2011 | accuracy: 77.9 -> 81.9 | Improved |
| fine-tune B-CNN (VGG-D) end to end | B-CNN (VGG-D) · Stanford Cars | accuracy: 82.9 -> 90.6 | Improved |
| replace indirectly fine-tuned FV with directly fine-tuned NetFV | FV (VGG-D) · CUB-200-2011 | accuracy: 74.7 -> 79.9 | Improved |
| replace fully connected pooling with bilinear pooling (f1 augmentation) | VGG-M trained from scratch · ImageNet val | top-1 error: 46.4 -> 38.7 | Improved |
| add f25 translation jitter to FC pooling (ablation) | VGG-M FC · ImageNet val | top-1 error: 46.4 -> 39.6 | Improved |
| add f25 jitter to bilinear pooling (ablation) | B-CNN VGG-M · ImageNet val | top-1 error: 38.7 -> 36.6 | Improved |
| add multi-scale features (ablation) | B-CNN (VGG-D) · DTD | mean per-class accuracy: 69.6 -> 72.9 | Improved |
| add multi-scale features (ablation) | B-CNN (VGG-D) · MIT indoor | mean per-class accuracy: 72.8 -> 79.0 | Improved |
| replace FV pooling with B-CNN pooling for texture | VGG-D relu5_3 features · DTD, FMD, KTH-T2b, MIT indoor | mean per-class accuracy: 73.6 vs 72.9 on DTD; 80.8 vs 81.6 on FMD | No clear effect |
| fine-tune B-CNN | B-CNN (VGG-D) · MIT indoor | accuracy: 72.8 -> 73.8 | Improved |
| project one feature with PCA to reduce bilinear dimension 16x (ablation) | B-CNN (VGG-M) · CUB-200-2011, FGVC-Aircraft | accuracy: less than 1% loss in accuracy with 32x512 | No clear effect |
| use NetVLAD with 32 components instead of projected B-CNN of the same size | NetVLAD · fine-grained datasets | accuracy: about 3-4% less accurate | Hurt |
Closest papers by meaning
- Contextual Convolutional Networks (ICLR 2023 · similarity 0.91)
- Involution: Inverting the Inherence of Convolution for Visual Recognition (CVPR 2021 · similarity 0.91)
- UniFormer: Unifying Convolution and Self-Attention for Visual Recognition (TPAMI 2023 · similarity 0.91)
- ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models (NeurIPS 2019 · similarity 0.91)
- AsCAN: Asymmetric Convolution-Attention Networks for Efficient Recognition and Generation (NeurIPS 2024 · similarity 0.90)
- SparseFormer: Sparse Visual Recognition via Limited Latent Tokens (ICLR 2024 · 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.
