General Object Detection

General Object Detection

1. Object Recognition

   • Segmentation
   • Scene Understanding
   • Object Tracking
   • Image Captioning
   • Event Detection
   • Activity recognition
   • Finding and Identifying objects of inerest

2. Object Detection

   • Artificial Intelligence
   • Robot Vision
   • Security
   • Human Computer Interaction
   • Detection of Specific Instances
     • Humans
     • Cars
     • Bicycles
   • Detection of Specific Categories
     • The face of a person
     • Detecting Instances of a paricular objects
     • building in an image
   • Find the Bounding Box for each Object
   • Object Detection Framework
     • Cascading
     • Sharing Feature Computation
     • Reduction per-window computation
     • Two Stage Detection
       • RCNN
         • Region Proposal
           • Crope and resize the Regions
             • SVM is trained using features from CNN and this replaces the Sofmax classifier
               • Bounding Box regression is learned for each object class with CNN features
             • Finetune an already trained network like ImageNet
         • RCNN Can't use any size Image Becaus of FC
           • SPPNet
             • After a Classic Convolutional Layer
               • Spatial Pyramid Pooling (SPP)
                 • Significant Speed up
                 • No Sacrifice of Quality
                 • Limited Accurace in very deep Network
             • Improve Detection Speed and Quality
               • Fast RCNN
                 • End-to-end training
                   • learns a softmax classifier and bounding box using a multitask loss
                     • No Softmax Classifier, SVM, BBRs in three seperate stages like RCNN and SPPnet
                     • Shares the Computation of Convolution across region proposals
                       • Adds Region of Interest pooling layer between the last CONV layer and first FC layer
                         • After FC the final Branvh has two output layers
                           • Softmax Propabilities for Object Category Prediction
                           • Class specific Bounding Box regression offsets for proposal refinment
                 • 3 times faster training
                 • 10 times faster testing
                 • Higher detection quality
                 • Single stage training process that updates all network layers
                 • No storage required for feature caching
                 • Relies of External region proposal
                 • Region Proposal is happening from neural network
                   • Faster RCN
                     • Region Proposal Network (RPN)
                     • First initializes k reference boxes( anchors) size (n,n)
                       • Each of them mapped to a lower dimensional vector
                         • Fed into two sibling FC layers
                           • Object Category Classification
                           • Box Regression Layer
                     • RFCN (Region Based Fully Convolutional Network
                       • Extract features with CONV
                         • RPN CONV and then two
                           • Object Classification
                             • Multiclass classification for each RoI
                           • Bounding Boc Regression
                             • Bounding Box Regression for Each RoI
                       • Mask RFCN
                         • Basically Ads a Branch which outputs a binary mask for each RoI
                           • It is a Fully Convolutional Network (FCN)
                             • Simple Train
                             • Generalizes well
                             • Adds only a small overhead to Faster RCNN
                             • Running at 5FPS
                         • Light Head RCNN
                           • Further Speed up to detection speed of RFCN
                             • Making the head of the detection network as light as possible to reduce the RoI regression computation
                             • Apply a large kernel seperable convolution to
     • One Stage Detection

3. Object Detection Competition

   • PASSCAL VOC
   • ILSVRC
   • MS COCO

4. Object Classification

   • Assesment of a Class in an Object

5. Two Competing Goals

   • High Quality/accuracy
     • Accurately Localinze
     • Accurately Recognize
     • Vast Range of Intercalss Variations
       • Instrisic Factors
         • Color
         • Tecture
         • Shape
       • Image Conditions
         • Weather
         • Background
         • Illuminations
     • Other Challenges
       • Noise Corruption
       • Poor Resolution
       • Filtering Sidtorton
   • High Efficiency
     • High Frame Rate
     • Acceptable Memory
     • Acceptable Storage
     • Exponential Increate in the number of Images
     • localize and recognize instances of very large number of Object
     • Scalability (detect unseen objects)

6. Deep Learning have been developed

   • GPU
   • Availability of large scale datasets
   • ImageNet, MS COCO