(1) speed/accuracy trade-offs for modern convolutional object detectors
Its main consideration is three kinds of detectors (Faster RCNN,R-FCN,SSD) as the meta structure, three kinds of CNN Network (vgg,inception,resnet) as feature extractor, change other parameters such as image resolution, proposals quantity, etc. The tradeoff between accuracy rate and speed of target detection system is studied.
(2) Yolo9000:better, Faster, stronger
It is an upgraded version of YOLO, which has made some improvements to the YOLO method, better considerations (Batch normalization, High Resolution, Anchor Boxes, Dimension Clusters, Direct Location prediction, fine-grained Features, Multi-scale Training), faster proposed a new Darknet-19 structure. In addition, it proposes a method of joint training on target detection data set and image classification data set, and constructs a hierarchical tree structure wordtree of visual content.
(3) A-fast-rcnn:hard positive generation via adversary for object detection
The main consideration is to learn a target detector with invariance for occlusion and deformation, a data-driven strategy for collecting large data sets with various state targets, such as occlusion and deformation subject to long tail theory, so as to learn a challenge network (ASTN,ASDN) to generate difficult samples with occlusion and deformation, Through the game between generator and detector, the target detector can improve the overall target detection performance of various state objects.
(4) Feature Pyramid Networks for Object Detection
The main consideration is to construct the feature pyramid by using the multi-scale pyramid structure inherent in the deep convolutional network, and create a top-down architecture with a transverse connection to construct advanced semantic feature maps on all scales, FPN as a general feature extractor, and the paper studies show that Although deep convnets have strong expressive ability and inherent robustness to scale change, it is still important to use pyramid representation to solve multiscale problems.
(5) Ron:reverse Connection with objectness Prior Networks for Object Detection
It mainly considers two problems, multi-scale target location and negative sample mining, and puts forward reverse connection so that the network can reduce the searching space of the target and objectness prior in the multi-scale detection target of CNN, and finally through the multi-task loss joint optimization reverse Connection, objectness Prior and target detectors.
(6) Accurate single Stage Detector Using recurrent rolling convolution
This paper proposes a new end-to-end training target detection network, which introduces recurrent rolling convolution structure to create "deep in context" classifier and regression in multi-scale feature map, and it mainly considers "top-down/bottom-up" Feature integration.
(7) Mimicking Very efficient Network for Object Detection
The current target detector needs to be initialized from the pre-trained imagenet classification model, which can achieve better results than training from zero, while the pre-trained image classification model is not optimal for the detection task, it mainly considers training high-efficiency detector without imagenet training. Its research has a network that satisfies the detection performance how to instruct the training of other networks, that is, using one detection network to supervise another more efficient network and maintain the accuracy rate, it proposes the characteristic mimic technology.
(8) Perceptual generative adversarial Networks for Small Object Detection
Small targets are difficult to detect because of their low resolution and noise, and the existing methods mainly consider the feature representation of all targets in Multiscale learning, which is limited by computational complexity. The main consideration is to establish a single architecture to address the small target detection problem, which promotes the presentation of small targets to "super-resolved", thus achieving features similar to large targets, making the detection task more discerning. The perceptual Gan model is proposed by using the generated anti-network, and the small target detection is improved by narrowing the difference between small target and large target.