学习日记2020-11-20论文学习

Highly Efficient Salient Object Detection with 100K Parameters

1. a flexible convolutional module, namely gOctConv, to efficiently utilize both in-stage and cross-stages multi-scale features for SOD task, while reducing the representation redundancy by a novel dynamic weight decay scheme.
2. we build an extremely light-weighted SOD model, namely CSNet–achieves comparable performance with ∼ 0.2% parameters
   
   
   
   
   

Dynamic Feature Integration for Simultaneous Detection of Salient Object, Edge and Skeleton

1.a dynamic feature integration strategy to explore the feature combinations automatically according to each input and task, and solve three contrasting tasks simultaneously in an end-to-end unified framework

Progressive Feature Polishing Network for Salient Object Detection

1. a novel multi-level representation refinement method for salient object detection, as well as a simple and tidy framework PFPN to progressively polish the features in a recurrent manner.
   2.For each polishing step, we propose the FPM to refine the representations which preserves the dimensions and hierarchical structure of the feature maps. It integrates highlevel semantic information directly to all low level features to avoid the long-term dependency problem.
   
   
   

F3Net: Fusion, Feedback and Focus for Salient Object Detection

1.the cross feature module-- fuse features of different levels, which is able to extract the shared parts between features and suppress each other’s background noises and complement each other’s missing parts.
2.the cascaded feedback decoder–can feedback features of both high resolutions and high semantics to previous ones to correct and refine them for better saliency maps generation.
3.pixel position aware loss–assign different weights to different positions

CapSal: Leveraging Captioning to Boost Semantics for Salient Object Detection

Specifically, we develop a CapSal model which consists of two sub-networks, the ImageCaptioning Network (ICN)
and the Local-Global Percep-tionNetwork(LGPN).

1.the usefulness of captioning for salient object detection
2.CapSal–to leverage the captioning information together with the local and global visual contexts for predicting salient regions
3.establish a COCO-CapSal dataset

Gradient-Induced Co-Saliency Detection

1.a gradient-induced co-saliency detection (GICD) network for Co-SOD.
2.a gradient inducing module(GIM)–to the discriminative co-salient features
3.attention retaining module (ARM)-- to keep the attention during the top-down decoding
4.a jigsaw strategy to train Co-SOD models
5.a challenging CoCA dataset with meticulous annotations

Global Context-Aware Progressive Aggregation Network for Salient Object Detection

1.Global Context-Aware Progressive Aggregation Network (GCPANet)
2.Feature Interweaved Aggregation (FIA) module---- integrates the low-level detail information, high-level semantic information, and global context information in an interweaved way
3.Self Refinement(SR)module
4.HeadAttention(HA)module
5.Global Context Flow (GCF) module

Re-thinking Co-Salient Object Detection

1.First, we construct a challenging CoSOD3k dataset, with more realistic settings.
2. CoEG-Net
3. provided a comprehensivestudy by summarizing 34 cutting-edge algorithms, benchmarking 16 of them over two classical datasets

CoADNet: Collaborative Aggregation-and-Distribution Networks for Co-Salient Object Detection

1. CoADNet provides some insights and improvements in terms of modeling and exploiting inter-image relationships in the CoSOD workflow
   2.an online intra-saliency guidance (OIaSG) module for supplying saliency priorknowledge
   3.a two-stage aggregate-and-distribute architecture to learn group-wise correspondences and co-saliency features – a group-attentional semantic aggregation (GASA) module – long-range semantic dependencies
   a gated group distribution (GGD) module-- distribute the learned group semantics to different individuals in a dynamic and unique way
   4.A group consistency preserving decoder (GCPD) is designed to replace conventional up-sampling or deconvolution driven feature decoding structures
   
   
   
   
   

n-Reference Transfer Learning for Saliency Prediction

1. a model-agnostic few-shot transfer learning paradigm to transfer knowledge from the source domain to the target domain.
2. a n-reference transfer learning framework to adaptively guide the training process