Keyword [Pairwise Ranking]
Li Y, Song Y, Luo J. Improving pairwise ranking for multi-label image classification[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 3617-3625.8.
1. Overview
1.1. Motivation
- hinge loss is non-smooth and difficult to optimize
- simple heuristics (top-k, threshhold) limits the use in the real world
In this paper,
- propose smooth pairwise ranking loss
- incorporate label decision into model
1.2. Related Work
1.2.1. Objective
- exact match
- Y^. predicted labels of i-th sample
- Y. GT
Hamming distance
Ranking objective
2. Algorithm
- f. label prediction, d–>K (K. all labels)
- g. label decision, K–>k (k<K)
2.1. Label Prediction
2.1.1. Pairwise Loss (PWE)
2.1.2. Log-Sum-Exp
- sample at most t pairs
2.2. Label Decision
- estimate label count
- estimate optimal thresholds for each class
- g. MLP on top of f’(x) (final CNN layer) (FC-ReLU, FC-ReLU) + two different branch
2.2.1. Label Count Estimation
- n-way classification
- Cross Entropy
2.2.2. Threshold Estimation
2.3. Details
- label prediction model
- VGG16 replace Softmax loss with LSEP loss
- finetune 10 epoches
- label decision model
- maximum number count = 4
- first FC-100, second FC-10
- count FC-4, threshold FC-14