aiacademy: 深度學習 CNN Object Recognition and Detection

###　Fine-grained object recognition

• CNN-­based computer vision applicatons
  • Find-grained object recognition
  • Object detection
  • Semantic segmentation
  • Image super resolution
  • Image style transfer
  • Action and gesture recognition
  • Image matching and co-segementation

• Introduction

  • Generic and fine-grained visual recognition

    • A large class number
    • Large intra-class
    • Subtle inter-class variations

    

• Part-based Method

• Idea

  • Neuron: part detector
  • Feature map: the spatial occurrence of certain part

  

We introduce the co-occurrence laye!r to encode the interaction between object parts.

Apporach (Co-occurrence layer)

Experiments: Dataset

• experiments: setting

  

• Visualization

---

Object Detection

• CNN-­based computer vision applicatons
  • Find-grained object recognition
  • Object detection
  • Semantic segmentation
  • Image super resolution
  • Image style transfer
  • Action and gesture recognition
  • Image matching and co-segementation

• Object Detection

  • Goal: Detecting instances of semantic objects of certain classes
  • Critical to high-level vision tasks such as surveillance, self-driving car, and image retrieval

  

R-CNN: Regions with CNN Features

Object Detection + CNN ===> R-CNN 第一個用　CNN 來做　Object Detection 的論文！

• Proposal extraction: Using selective search [Uijlin et al.,IJCV’13]
• Compute CNN features in the layer ‘fc7’ of Caffe CNN
• Region classification: linear SVMs or a softmax classifer
• Regression-based bounding box refinement

Fast R-CNN

• Apply fully concolutional networks to the whole image
• Rol pooling: each proposal is pooled into a fix-size feature map
• Classification with a softmax layer
• Regression-based bounding box refinement

Experimental Results

Faster R-CNN

R-CNN vs. Fast-R-CNN vs. Faster R-CNN

YOLO9000 [Redmon & Farhadi, CVPR’17]