基于深度神经网络的目标检测概要.pptx

 第1页 | 共25页 基于深度神经网络的目标检测 检测 ≈ 定位 + 分类 第2页 | 共25页 第3页 | 共25页 传统目标检测 基于候选区域的目标检测 回归方法的深度学习目标检测 YOLO 第4页 | 共25页 滑动窗口策略 手工设计特征 SVM、adaboost等 两个问题 传统目标检测 第5页 | 共25页 候选区域selective search IOU NMS 非极大值抑制 PASCAL VOC上的检测率从35.1%提升到53.7% 基于候选区域的目标检测 第6页 | 共25页 Step1. Input an image Step2. Use selective search to obtain ~2k proposals Step3. Warp each proposal and apply CNN to extract its features Step4. Adopt class-specified SVM to score each proposal Step5. Rank the proposals and use NMS to get the bboxes. Step6. Use class-specified regressors to refine the bboxes’ positions. TOO SLOWWWW !!! － SPPNET RCNN三个问题 （分阶段训练、空间浪费、慢47s） SPP-Net: Motivation Cropping may loss some information about the object Warpping may change the object’s appearance 第7页 | 共25页 FC layer need a fixed-length input while conv layer can be adapted to arbitrary input size. Thus we need a bridge between the conv and FC layer. Here comes the SPP layer. 第8页 | 共25页 SPP-Net: Training for Detection(1) 第9页 | 共25页 Conv5 feature map Conv5 feature map Conv5 feature map Image Pyramid FeatMap Pyramids conv Step1. Generate a image pyramid and exact the conv FeatMap of the whole image SPP-Net: Training for Detection(2) Step 2, For each proposal, walking the image pyramid and find a project version that has a number of pixels closest to 224x224. (For scaling invariance in training.) Step 3, find the corresponding FeatMap in Conv5 and use SPP layer to pool it to a fix size. Step 4, While getting all the proposals’ feature, fine-tune the FC layer only. Step 5, Train the class-specified SVM 第10页 | 共25页 SPP-Net: Testing for Detection Almost the same as R-CNN, except Step3. 第11页 | 共25页 第12页 | 共25页 Speed: 64x faster than R-CNN using one scale, and 24x faster using five-scale paramid. mAP: +1.2 mAP vs R-CNN 2. 训练花费过大的硬盘开销和时间 1. 训练分多阶段，并不是端到端的训练过程 FC layers Conv layers SVM regressor store 第13页 | 共25页 3. 训练sppnet只微调全连阶层（检测除了语义信息还需要位置信息，多层pooling操作导致位置信息模糊） Fast R-CNN Fast R-CNN: Motivation Ross Girshick, Fast R-CNN, Arxiv tec