------------------------------------------------------------------------
About Face detection
------------------------------------------------------------------------
1、Finding Tiny Faces
Code:https://github.com/peiyunh/tiny
小目標檢測難3大原因:目標本身尺度變化、圖像分辨率以及環境因素。本文針對多尺度訓練了不同的檢測器,這些檢測器所用特征來自同一網絡的不同層級。此外,還充分利用了目標周邊信息。
2、Seeing Small Faces from Robust Anchor’s Perspective
基於anchor設計原理解決小臉檢測不到的問題。
3、Face-MagNet: Magnifying Feature Maps to Detect Small Faces
Code:https://github.com/po0ya/face-magnet
基於Faster-RCNN框架提出Face-MagNet網絡(在人臉建議和分類前放大特征圖的判別能力)而無需任何跳過或殘差連接。在RPN中和ROI前都加了一組反卷積層。另外,評估了其他3種針對尺度問題而有較好調整架構的方法:context pooling, skip connections, and scale partitioning.
4、Detecting and counting tiny faces
Code:https://github.com/alexattia/ExtendedTinyFaces
對Finding Tiny Faces這篇文章的深入理解,類似的方法。
5、SSH: Single Stage Headless Face Detector
Code:https://github.com/mahyarnajibi/SSH
單階段檢測器,速度快,占用內存少,在不同深度的網絡層進行人臉檢測,用於檢測大、中、小人臉。
6、S3FD: Single Shot Scale-invariant Face Detector
Code:https://github.com/sfzhang15/SFD
(1) proposing a scale-equitable face detection framework to handle different scales of faces well.
(2) improving the recall rate of small faces by a scale compensation anchor matching strategy.
(3) reducing the false positive rate of small faces via a max-out background label.
7、Multi-Path Region-Based Convolutional Neural Network for Accurate Detection of Unconstrained “Hard Faces”
a two-stage cascaded face detection framework:
(1) a Multi-Path Region Proposal Network(MP-RPN),利用3個平行特征圖的輸出預測不同尺度的候選人臉區域,嵌有帶有上采樣過濾的卷積層和新提出的產生“難”例采樣層。
(2) a Boosted Forests classifier,利用候選人臉區域內的深層面部特征和周圍更大區域的上下文特征,大大減少 hard negative samples.
8、Scale-Aware Face Detection
先對圖片上的人臉進行尺度估計,再在特定尺度上進行人臉檢測(使用RPN,只使用一種anchor,且每次只檢測一張臉)。不用在各個尺度下對人臉檢測,因此速度較快。
9、Detecting Faces Using Inside Cascaded Contextual CNN
不是使用多個CNN網絡來級聯的,而是使用一個CNN中不同網絡層來做級聯。CNN網絡的前幾層完成簡單的人臉檢測,后面的網絡完成難度較大的人臉檢測,采用data routing機制來使不同的卷積層由不同類型的樣本來訓練,關注於去掉不同類型的非人臉樣本。 同時使用 body part localization 來輔助人臉檢測。
10、Face Detection through Scale-Friendly Deep Convolutional Networks
核心方法類似SSD。在網絡不同階段引出分支檢測對應范圍的人臉。訓練時針對不同分組只用對應尺度的樣本進行訓練。
11、A Multi-Scale Cascade Fully Convolutional Network Face Detector
基於FCNs的3層級聯結構。It first proposes the approximate locations where the faces may be, then aims to find the accurate location by zooming on to the faces. Each level of the FCN cascade is a multi-scale fully-convolutional network, which generates scores at different locations and in different scales. A score map is generated after each FCN stage. Probable regions of face are selected and fed to the next stage. The number of proposals is decreased after each level, and the areas of regions are decreased to more precisely fit the face.
12、Face Detection using Deep Learning: An Improved Faster RCNN Approach
對Faster RCNN的一些改進策略: feature concatenation, hard negative mining, multi-scale training, model pretraining, and proper calibration of key parameters.
13、Face R-CNN
對Faster RCNN改進:new multi-task loss function design, online hard example mining, and multi-scale training strategy
14、Face Detection Using Improved Faster RCNN
multi-scale training, multi-scale testing, light-designed RCNN, keep the small proposals at training and testing stage, directly select top-ranked proposals (e.g., 6000) without NMS in the RPN stage for R-CNN, a vote-based NMS ensemble strategy.
15、Anchor Cascade for Efficient Face Detection
propose a context pyramid maxout mechanism for anchor cascade。大大減少計算量和提高檢測精度。同時對於訓練小規模模型也有很高的檢測精度。
16、SFace: An Efficient Network for Face Detection in Large Scale Variations
解決大尺度變化問題。提出新算法SFace:整合了anchor-based methods(類似RetinaNet)和anchor-free based methods(類似UnitBox)。
-----------------------------------------------------------------------
-----------------------------------------------------------------------
1、Single-Shot Refinement Neural Network for Object Detection
Code:https://github.com/sfzhang15/RefineDet
可看做將Faster RCNN的two stages檢測方法和SSD結合。
propose a novel one-stage framework(RefineDet) consists of two inter-connected modules. the former aims to (1) filter out negative anchors to reduce search space for the classifier, and (2) coarsely adjust the locations and sizes of anchors to provide better initialization for the subsequent regressor. The latter module takes the refined anchors as the input from the former to further improve the regression and predict multi-class label. Meanwhile, we design a transfer connection block to transfer the features in the anchor refinement module to predict locations, sizes and class labels of objects in the object detection module. The multitask loss function enables us to train the whole network in an end-to-end way.
2、An Analysis of Scale Invariance in Object Detection-SNIP
可看成改版版本的Image Pyramid。
分析了小尺度與預訓練模型尺度之間的關系, 提出了Scale Normalization for Image Pyramids (SNIP):在訓練中,每次只回傳那些大小在一個預先指定范圍內的proposal的gradient,而忽略掉過大或者過小的proposal;在測試中,建立大小不同的Image Pyramid,在每張圖上都運行這樣一個detector,同樣只保留那些大小在指定范圍之內的輸出結果,最終在一起NMS。這樣就可以保證網絡總是在同樣scale的物體上訓練,也就是標題中Scale Normalized的意思。
3、Cascade R-CNN: Delving into High Quality Object Detection
Code:https://github.com/zhaoweicai/cascade-rcnn
基於two-stage detector。Cascade R-CNN是R-CNN的multi-stage延伸,由一系列隨着IOU臨界值增加而訓練的檢測器構成,從而對close false positives更具有選擇性。R-CNN階段的cascade是按順序訓練的,使用一個階段的輸出來訓練下一階段。類似於boostrapping methods,不同點是Cascade R-CNN的重采樣過程並不旨在mine hard negatives,而是通過調整bounding boxes,每個階段的目的都是為了找到一組好的false positive來訓練下一階段。
4、Single-Shot Object Detection with Enriched Semantics
在SSD網絡基礎上,增加了語義分割分支和全局激活模塊。前者增加低層檢測特征,后者通過學習特征通道和目標類別的語義關系來進行高層目標檢測特征。
5、Multi-scale Location-aware Kernel Representation for Object Detection
Code:https://github.com/Hwang64/MLKP
提出了一種新穎的多尺度位置感知核表示(MLKP),將判別性高階統計量結合到object proposals的表示中以進行有效的對象檢測。MLKP基於多項式核近似,可以有效生成低維高階表示,其固有的位置保持性和敏感性也保證了可以靈活地用於目標檢測。
6、A-Fast-RCNN: Hard Positive Generation via Adversary for Object Detection
Code:https://github.com/xiaolonw/adversarial-frcnn
提出學習一個可以生成遮擋和變形樣本的對抗網絡,對抗器的目標是生成讓目標檢測器難以進行分類的樣本。在我們的框架中,原始檢測器和對抗器都是以聯合的方式學習的。
7、Detecting Small Signs from Large Images
large images are broken into small patches as input to a Small Object-Sensitive-CNN (SOS-CNN) modified from a Single Shot Multibox Detector (SSD) framework with a VGG-16 network as the base network to produce patch-level object detection results. Scale invariance is achieved by applying the SOS-CNN on an image pyramid. Then, image-level object detection is obtained by projecting all the patch-level detection results to the image at the original scale.
8、Perceptual Generative Adversarial Networks for Small Object Detection
P-GAN將小目標的特征映射到相似的大目標特征上來縮小差別,便能將小目標足夠近似到大目標來欺騙判別器,達到小目標檢測的目的。
9、Feature Pyramid Networks for Object Detection
特征金字塔網絡。
10、SOD-MTGAN: Small Object Detection via Multi-Task Generative Adversarial Network
提出一個對於小目標檢測的標准的端到端的多任務生成對抗網絡(MTGAN),適用於任何已有的檢測器。In the MTGAN, the generator network produces super-resolved images and the multi-task discriminator network is introduced to distinguish the real high-resolution images from fake ones, predict object categories, and refine bounding boxes, simultaneously. More importantly, the classification and regression losses are back-propagated to further guide the generator network to produce super-resolved images for easier classification and better localization.
11、Deep Feature Pyramid Reconfiguration for Object Detection
當前特征金字塔的設計在如何整合不同尺度的語義信息方面仍然不夠高效。本文把特征金字塔轉換為特征的再組合過程,創造性地提出了一種高度非線性但是計算快速的結構將底層表示和高層語義特征進行整合。該網絡由兩個模塊組成:全局注意力和局部再組合。這兩個模塊分布能全局和局部地去在不同的空間和尺度上提取任務相關的特征。重要的是,這兩個模塊具有輕量級、可嵌入和可端到端訓練的優點。
12、Parallel Feature Pyramid Network for Object Detection
使用SPP模塊通過擴大網絡寬度而不是增加深度來生成金字塔形特征圖。提出MSCA模塊有效地組合了不同規模的上下文信息。
13、SAN: Learning Relationship between Convolutional Features for Multi-Scale Object Detection
提出了Scale Aware Network (SAN),將來自不同尺度的卷積特征映射到尺度不變的子空間,並設計了一種獨特的學習方法,純粹考慮了沒有空間信息的渠道之間的關系。所提出的SAN減少了標度空間中的特征差異並提高了檢測精度。
14、A CLOSER LOOK: SMALL OBJECT DETECTION IN FASTER R-CNN
介紹了一種生成anchor proposals的改進建議,並對Faster R-CNN進行修改,利用較高分辨率的小目標的feature maps。
15、Improving Small Object Proposals for Company Logo Detection
we introduce an improved scheme for generating anchor proposals and propose a modification to Faster R-CNN which leverages higher-resolution feature maps for small objects.
16、Scale-aware Pixel-wise Object Proposal Networks
提出Scale-aware Pixel-wise Object Proposal(SPOP)網絡,可以生成具有高召回率和平均最佳重疊(ABO)的proposals,即使對於小目標也是如此。另外,引入了一個類似分段的像素定位網絡來密集預測每個像素的對象坐標,並開發了一種尺度感知對象定位策略,該策略將來自大尺寸和小尺寸網絡的預測與加權機制相結合,以提高各種對象尺寸的坐標預測精度。
———————————————————————————————
原文鏈接:https://blog.csdn.net/u014236392/article/details/83993730