05_Convolutional_Neural_Networks_part_2.md

#5. Convolutional Neural Networks part 2

Limitations about trational CNN:

• Mostly on centered images
• Only a single object per image
• Not enough for many real world vision tasks:
  • Localisation
  • Object Detection
  • Semantic segmentation
  • Instance segmentation

Localisation

• Single object per image
• Predict coordinates of a bounding box (x,y,w,h)
• Evaluate via IoU

Object detection

We don't know the number of objects in the image.

• Object proposal : find region of interest (ROIs) in the image
• Object classification : classify the object in these regions

Two main families:

• A grid in the image where each cell is a proposal (SSD, YOLO)
• Region proposal (SPP, MultiBox, Faster RCNN)

YOLO

SDD : single-shot detector

Box proposals

Instead of having a predefined set of box proposals, find them on the image:

• Selective search: from pixels (not learnt)
• Faster-RCNN : region proposal network(RPN)

Crop-and-resize operator (Rol-Pooling)

Fast-RCNN

Faster-RCNN

State of the art

• comparison between different algo
• Mask RCNN, light-head R-CNN for best accuracy
• Yolo, SSD, Light-Head R-CNN for fast inference

Segmentation

Mask R-CNN

State of art

• Mask-RCNN and other architectures
• Focal loss, Feature Pyramid Networks, etc
• Retina Net
• MegDet