Keyword [Multi-Scale]
Cheng G, Gao D, Liu Y, et al. Multi-scale and Discriminative Part Detectors Based Features for Multi-label Image Classification[C]//IJCAI. 2018: 649-655.
1. Overview
1.1. Motivation
- global CNN feature lack geometric invariance for addressing the problem of intra-class variation
In this paper, it proposes Multi-scale and Discriminative Part Detectors (MsDPD)
- task-driven feature pooling
1.2. Architecture
1.3. Formulation
- S = softmax(); sigma = sigmoid()
- xij. a patch of origin image
- phi. pooled DPD-based feature
- O(xij). a point of DPD-based feature map (1x1xK)
- P. prediction
1.4. Loss Function
classification loss. BCE
generalized max pooling regularization term
- enforce the dot product similarity between O(xij) and the pooled feature phi(Xi) to be a constant one
- object part-level classification loss term
- y_l∈R^K. stands for object part label vector of object part instance xl