Research map / Detection and segmentation / Open-set and few-shot detection
Fine-Grained Object Classification via Self-Supervised Pose Alignment
Check 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 |
|---|---|---|---|
| Use P2P-Net with ResNet50 instead of the strongest reported benchmark baseline | Fine-grained classification · CUB | accuracy (%): 90.0 -> 90.2 | Helped |
| Use P2P-Net with ResNet50 instead of the strongest reported benchmark baseline | Fine-grained classification · CAR | accuracy (%): 95.3 -> 95.4 | Helped |
| Use P2P-Net with ResNet50 instead of the strongest reported benchmark baseline | Fine-grained classification · AIR | accuracy (%): 93.9 -> 94.2 | Helped |
| Add curriculum supervision (ablation) | ResNet50 baseline · CUB, CAR, AIR | accuracy (%): CUB 85.5 -> 88.4; CAR 92.7 -> 94.9; AIR 90.3 -> 93.8 | Helped |
| Use label smoothing instead of one-hot labels (ablation) | P2P-Net curriculum supervision · CUB | accuracy (%): 81.7, 87.3, 85.8, 87.8, 88.4; one-hot: 82.1, 87.0, 85.6, 87.3, 88.2 | Mixed |
| Add feature regularization without unsupervised part alignment (ablation) | ResNet50 baseline · CUB, CAR, AIR | accuracy (%): CUB 85.5 -> 89.0; CAR 92.7 -> 94.8; AIR 90.3 -> 92.0 | Mixed |
| Add unsupervised part alignment to feature regularization (ablation) | Baseline+FR · CUB, CAR, AIR | accuracy (%): CUB 89.0 -> 89.0; CAR 94.8 -> 95.0; AIR 92.0 -> 92.5 | Mixed |
| Add unsupervised part alignment to curriculum supervision and feature regularization (ablation) | Baseline+CS+FR · CUB, CAR, AIR | accuracy (%): CUB 90.0 -> 90.2; CAR 95.0 -> 95.4; AIR 93.9 -> 94.2 | Helped |
| Concatenate local-part and global-image features instead of using feature regularization (ablation) | ResNet50 baseline · CUB, CAR, AIR | accuracy (%): CUB: 88.4 vs 89.0; CAR: 94.7 vs 94.8; AIR: 93.8 vs 92.0 | Mixed |
| Use feature regularization with unsupervised part alignment rather than feature concatenation (ablation) | ResNet50 representation learning · CUB, CAR, AIR | RMSE: CUB 0.268 -> 0.213; CAR 0.213 -> 0.179; AIR 0.354 -> 0.263 | Helped |
| Vary the number of detected parts (ablation) | P2P-Net · CUB | accuracy (%): N=2: 88.1; N=3: 89.9; N=4: 90.2; N=5: 90.2; N=6: 90.0 | Mixed |
Closest papers by meaning
- Revisiting Pose-Normalization for Fine-Grained Few-Shot Recognition (CVPR 2020 · similarity 0.87)
- Part-Stacked CNN for Fine-Grained Visual Categorization (CVPR 2016 · similarity 0.87)
- Interpretable and Accurate Fine-grained Recognition via Region Grouping (CVPR 2020 · similarity 0.86)
- Weakly Supervised Learning of Part Selection Model with Spatial Constraints for Fine-Grained Image Classification (AAAI 2017 · similarity 0.86)
- Dual Pose-invariant Embeddings: Learning Category and Object-specific Discriminative Representations for Recognition and Retrieval (CVPR 2024 · similarity 0.84)
- Democratizing Fine-grained Visual Recognition with Large Language Models (ICLR 2024 · similarity 0.83)
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.