A Simple Pooling-Based Design for Real-Time Salient Object Detection （2019 CVPR）
Global Guidance Module (GGM) : PPM feature 與 FPN 每個level 融合。增加RF，緩解obj預測不完整。
Feature Aggregation Module (FAM): 類似PPM，add代替concate。
1、reducing the aliasing effect of upsampling。
2、allows each spatial location to view the local context at different scale spaces, context at different scale spaces
joint training：edge data（BCE）和 salient data （CE）交替訓練，細化edge。
Contrast Prior and Fluid Pyramid Integration for RGBD Salient Object Detection （2019 CVPR）
depth label 與 SOD 結合。
Contrast-enhanced Net(CEN)：增加depth map前景背景的contrast，類內緊湊，類間可分。
original depth map contrast not obvious enough
cross-modal fusion：Enhanced map 作爲 attention map 
Fluid Pyramid Integration(FPI)：FPI leads all the highlevel features into low-level features for every node.
B+C>B+C+M, meets feature compatibility problem
LOSS：saliency cross-entropy loss + 5 contrast loss
