Improving Pairwise Ranking for Multi-Label Image Classification

# Summary

1. LSEP Loss (log-sum-exp pairwise)
2. Label Decision (Label count estimation + Threshold estimation)

# Difference from Paper

1. VGG16 -> Inception ResNet v2
2. 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