学习日记 2020-11-17

Optimizing the F-measure for Threshold-free Salient Object Detection

1.F-measure formulation in both training and evaluation for SOD
2.FLoss

F-measure ：
f-measure是一种统计量，F-Measure又称为F-Score，F-Measure是Precision和Recall加权调和平均，是IR（信息检索）领域的常用的一个评价标准，常用于评价分类模型的好坏。
在f-measure函数中，当参数α=1时，F1综合了P和R的结果，当F1较高时则能说明试验方法比较有效。

Stacked Cross Refinement Network for Edge-Aware Salient Object Detection

1.Cross Refinement Unit(CRU)
2.a novel framework：Stacked Cross Refinement Network(SCRN)

Towards High-Resolution Salient Object Detection

1.consists of a Global Semantic Network (GSN), a Local Refinement Network (LRN) and a Global-Local Fusion Network(GLFN)
2.the first high-resolution data set (HRSOD)

EGNet: Edge Guidance Network for Salient Object Detection

1.EGNet to explicitly model complementary salient object information and salient edge information
2. model allowing them to mutually help each other

Structured Modeling of Joint Deep Feature and Prediction Refinement for Salient Object Detection

1.a deep unified CRF saliency model —formulates messages passing with CRFs for joint feature and prediction refinement.
2. deep cascade CRFs architecture—seamlessly incorporated with CNN to integrate and refine multi-scale deep features and predictions for a refined saliency map

Employing Deep Part-Object Relationships for Salient Object Detection

1.a new property, i.e., part-object relationships by CapsNet
2.a deep TSPOANet for salient object detection adopts a two-stream strategy

DeepUSPS: Deep Robust Unsupervised Saliency Prediction With Self-Supervision

1. a novel systematic mechanism to refine the pseudo-ground-truth labels of handcrafted unsupervised saliency methods iteratively via self-supervision.