Keyword [Spatial Pooling] [Classwise Pooling]
Durand T , Mordan T , Thome N , et al. WILDCAT: Weakly Supervised Learning of Deep ConvNets for Image Classification, Pointwise Localization and Segmentation[C]// 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017.
1. Overview
In this paper, it proposes WILDCAT
- FCN backbone
- Multi-Map WSL transfer layer
- Wildcat Pooling
1.1. Related Work
- Max Pooling (MP)
- Global Average Pooling (GAP)
- LSE Pooling
- Compact Bilinear Pooling
2. Architecture
2.1. FCN
- ResNet101. output [W/32, H/32, 2048]
- remove GAP and FC
- replace with WSL transfer amd wildcat pooling layers
2.2. Multi-map Transfer Layer
- per class through 1x1 Conv
- (h, w, c) – (h, w, m*c)
2.3. Wildcat Pooling
2.3.1. Classwise Pooling
- (h, w, mc) – (h, w, c)
2.3.2. Spatial Pooling
- for a class map (h, w, 1)
- average k+ max point
- average k- max point
- α. trade off
- hypothesize that maximum scoring regions are more useful for classification
- With α < 1 Wildcat should focus more on discriminating regions and then better localize features than with α=1
2.4. Training
- input a single image
2.5. Inference
- classification. input a single image
- localization. extract the region with maximum score for each class
- segmentation. take the class with maximum score at eachlocation independently or apply CRF
3. Experiments
3.1. Details
- M = 4, α = 0.7
- image. 448x448