Tensorflow_Improving_Pairwise_Ranking_for_Multi-label_Image_Classification
Tensorflow_Improving_Pairwise_Ranking_for_Multi-label_Image_Classification copied to clipboard
shjo-april
[Re-implementation] Improving Pairwise Ranking for Multi-label Image Classification (CVPR2017)
Improving Pairwise Ranking for Multi-Label Image Classification
# Summary
- LSEP Loss (log-sum-exp pairwise)
- Label Decision (Label count estimation + Threshold estimation)
# Difference from Paper
- VGG16 -> Inception ResNet v2
- binary-cross-entropy (with sigmoid) -> Focal Loss
# Results (NUS-WIDE Tags : 1K)
| Model | Precision | Recall | mAP |
|---|---|---|---|
| WARP | 18.3% | 30.8% | 24.55% |
| CNN-RNN (CVPR2016) | 18.5% | 31.2% | 24.85% |
| S-CNN-RNN (CVPR2017) | 19.0% | 30.2% | 24.60% |
| Multi-label Triplet Embeddings (ICML 2018) | 19.8% | 32.7% | 26.25% |
| (self) Inception ResNet v2 (LSEP -> Label Decision) | 30.74% | 21.52% | 26.13% |
# Test Samples
# Tensorboard
1. LSEP Loss
2. Label Decision
# Reference
- WARP: Deep Convolutional Ranking for Multilabel Image Annotation
- CNN-RNN: A Unified Framework for Multi-label Image Classification
- Semantic Regularisation for Recurrent Image Annotation
- Improving Pairwise Ranking for Multi-label Image Classification
- Multi-label Triplet Embeddings for Image Annotation from User-Generated Tags
Metadata
20
Stars
5
Forks
Watchers
Owner
shjo-april
Metadata
[Re-implementation] Improving Pairwise Ranking for Multi-label Image Classification (CVPR2017)