[Tadeas] Image Detection

Object Detection

Bounding box(最小外截距项)和category label

Difference with the other task:

1. single task: classification <---> classification and localization
2. multi-task: object detection <---> instance segmentation

contest

imagenet large scale visual recognition challenge

URL: http://image-net.org/challenges/LSVRC/2017/index

• object detection

1. 2013-2017, 200 category, multi-category and bounding box each image
   

• other task

1. image classification
2. scene classification
3. object localization
4. scene parsing

PASCAL VOC

URL： http://host.robots.ox.ac.uk/pascal/VOC/

• eject Detection

  Bounding box(最小外截距项)和category label

  Difference with the other task:

1. single task: classification <---> classification and localization
2. multi-task: object detection <---> instance segmentation

contest

imagenet large scale visual recognition challenge

URL: http://image-net.org/challenges/LSVRC/2017/index

• object detection

1. 2013-2017, 200 category, multi-category and bounding box each image
   

• other task

1. image classification
2. scene classification
3. object localization
4. scene parsing

PASCAL VOC

URL： http://host.robots.ox.ac.uk/pascal/VOC/

• evolution：R-CNN --> SPP-Net --> Fast R-CNN --> Faster R-CNN --> YOLO --> SSD -->R-FCN

R-CNN

• model structure:

1. Region proposal: devisity boxing with different position and size
2. Classification: CNN classificator, heavy calculation
   
   
   $ $
   R-CNN model
   

$ $ R-CNN in models

• R-CNN trainning process

1. pre-train a imageNet CNN model --> Model1
2. fine-tune proposal regions with SS(selective search) -->Model2
   -- Log loss
   -- softmax layer --> (N+1) way
   -- 32 positive samples （N classes）: IoU with Ground truth > 0.5
   -- 96 negative samples (1 classes): IoU < 0.5

1. based on Model2, on its Fc7 train linear SVM classificator --> Model3
   -- Hinge loss
   -- each class with a SVM classficator (N classes)
   -- postive samples: all Ground-truth areas
   -- negative samples: IoU < 0.3 of SS of SS areas

1. with Model2, on Fc7, train Bounding box regression model
   -- improve bounding box accuracy
   -- a regression model for each class (N class)
   --- bounding box in SS regression：P --> G
   --- input:
   ----- bounding box \({P^i, G^i}_{i=1,...,N}\)
   ----- center: \((x, y)\), width and height:\((w, h)\)
   ----- $ P^i=(P^i_x, P^i_y, P^i_w, P^i_h), P^i=(P^i_x, P^i_y, P^i_w, P^i_h)$
   ----- feature for Conv5 of CNN: \(\phi_5(P)\)
   --- IoU for P > 0.6
   --- Squared loss:
   ----- \({{\rm{W}}_*} =\mathop {\arg \min }\limits_{{{{\rm{\hat w}}}_*}} {( {t_o^i - {{{\rm{\hat w}}}_o}{\phi_5}( {{P^i}} )} )^2} + \lambda ||{{\rm{\hat w}}_*}||^2\)
   ----- where \(t_x = (G_x-P_x)/{P_w}, t_y = (G_y-P_y)/{P_y}, t_w = log(G_w/{P_w}), t_h= log(G_h/{P_h})\)
   --- test process:
   ----- \({\hat G}_x = P_x d_x(P)+P_x, {\hat G}_y = P_y d_y(P)+P_y, {\hat G}_w = P_w exp(d_w(P)), d_*(P) = {\rm}^{\rm T}_* \phi_5 (P)\)

explain for cost function: Ref

1. R-CNN test
   -- SS (fast model) get about 2000 proposal areas/images
   -- resize(expand and scaling) them into size of 227*227
   -- Model2 to get 2 feature map: proposal areas subset and regressed Bbox
   --- for each class
   1). Fc7 feature --> SVM classificator --> class score
   2). with IoU > 0.5, get non-redundant area subset

• reorder all areas in decrease order with the score
• remove redundant areas: IoU > 0.5 with the area of highest score
• save the highest score area, the left areas to candidate set
  3). conv5 --> Bounding box regression model --> regressed Bbox
  4). with regressed Bboxs fine tune ares subset

1. performance evaluation

• mAp@5 (mean Average Precision)
  -- calculate AP for each class, then get mean value
• AP is the area of Precision-Recall Curve
• precision: TP/(TP+FP)
• recall: TP/(TP+FN)
  ** True positive area: IoU >=0.5 with ground Truth
  ** False positive area: IoU < 0.5
  ** False negative area: unmarked Gound truth area:
  ** IoU = Intersection over Unit = (AnB)/(AUB)

1. defect：

• too much time for training process
  -- fine-tune (18) + feature extraction (63) + SUM/Box training (3)
• too much time for test: a VGG16 47S
• complex and multi-training processes