文献阅读（96）目标检测综述（部分）

• 题目：Recent Advances in Deep Learning for Object Detection
• 时间：2019
• 网络：arXiv
• 研究机构：新加坡管理学院

1 Detection Components

• FPN: Feature Pyramid Networks特征金字塔

FPN combined deep layer features with shallow layer features to enable
object detection in feature maps at different scales

• YOLO的主要缺点：

1. it could detect upto only two objects at a given location, which made it difficult to
   detect small objects and crowded objects
2. only the last feature map was used for prediction, which was not suitable for predicting objects at multiple scales and aspect ratios.

• CornerNet: anchor-free framework

Class heatmaps calculated the probabilities of being corners, and corner offsets were used to regress the corner location

2 人脸识别

• 人脸识别和通用目标检测的不同之处在于：

1. the range of scale for objects in face detection is much larger than objects in generic detection
2. Face objects contain strong structural information, and there is only one target category in face detection

• 主要的数据集有

• WIDER FACE: 32203 images with about 400k faces for a large range of scales
• FDDB: 5171 faces in 2845 images
• PASCAL FACE: collected from PASCAL person layout test set, with 1335 labeled faces in 851 images
  Commonly face detectors will first be trained on a large scale dataset(WIDERFACE etc. ) and tested on FDDB

3 行人识别数据集

• CityPerson
• Caltech: one of the most popular and challenging datasets for pedestrian detection
  ETH: commonly it’s usedas test set to evaluate performanceof the models trained on the large scale datasets(CityPersons dataset etc.)
• INRIA
• KITTI: 7481 labeled images of resolution 1250x375 and another 7518 images for testing. The person class in KITTI is divided into two subclasses: pedestrian and
  cyclist