基于深度神经网络的目标检测.pptxVIP

基于深度神经网络的目标检测第页 | 共25页检测 ≈ 定位 + 分类第页 | 共25页YOLO第页 | 共25页传统目标检测滑动窗口策略手工设计特征SVM、adaboost等两个问题第页 | 共25页基于候选区域的目标检测候选区域selective searchIOUNMS非极大值抑制PASCAL VOC上的检测率从35.1%提升到53.7%第页 | 共25页RCNN三个问题（分阶段训练、空间浪费、慢47s）Step1. Input an imageStep2. Use selective search to obtain ~2k proposalsStep3. Warp each proposal and apply CNN to extract its featuresStep4. Adopt class-specified SVM to score each proposalStep5. Rank the proposals and use NMS to get the bboxes. Step6. Use class-specified regressors to refine the bboxes’ positions.TOO SLOWWWW !!! － SPPNET 第页 | 共25页SPP-Net: MotivationCropping may loss some information about the objectWarpping may change the object’s appearance第页 | 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.第页 | 共25页SPP-Net: Training for Detection(1)Step1. Generate a image pyramid and exact the conv FeatMap of the whole imageConv5 feature mapConv5 feature mapconvConv5 feature mapFeatMap PyramidsImage Pyramid第页 | 共25页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第页 | 共25页SPP-Net: Testing for DetectionAlmost the same as R-CNN, except Step3.第页 | 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第页 | 共25页1. 训练分多阶段，并不是端到端的训练过程Conv layersFC layersSVMregressor2. 训练花费过大的硬盘开销和时间store3. 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第页 | 共25页Fast R-CNN: MotivationJOINT TRAINING!!特征提取和分类放在一个网络之中，联合训练ROI pooling layer多任务损失函数(multi-task loss)Ross Girshick, Fast R-CNN, Arxiv tech report第页 | 共25页Fast R-CNN: Joint Training F