Weekend Reads - Advanced Models for Computer Vision

Key Notes

What Classifier will Miss - Human-level scene understanding

• Parsing the scene
• The angle of Bicycle (Pose, Relative pose)
• Person on Bicycle
• Closer Inspection

Tasks

• Object Detection
• Pose Estimation
• Accuracy vs Efficiency of Models

CNN as Deep Learning Puzzle

Input-Output Node, Loss Computation and Backprop

Classification - sparse description of the image

Object Detection

• Multi-task problem
• Classification & Localisation
• Object, Location, Bounding box
• Dataset, Samples, List of Objects, Labels, Bbox for each object

Predict BBOX Coordinates

• Continuous Output
• Minimize mse of samples
• Regression for bbox prediction
• The first part is the classification
• The Second Step is regression

Faster RCNN

• Two-Stage Detector
• Good Candidate BBOX
• Refine through Regression
• Discretize bbox space
• Anchor points distributed
• Candidate boxes of different scale and ratio
• n candidates per anchor
• Is there an object or not in the box
• Refine through regression
• We cannot backdrop on parameters of bbox (Spatial Transformer Networks)

One Stage Detector - Train end to end

• Employ Hard negative mining

Retinanet uses Focal Loss (The loss function is just a mathematical way of saying how far off a guess is from the real value of a data point.). It puts more weight on the objects that were hard to classify and decreases the impact on easy correct predictions

Semantic Segmentation

• Pooling - reduce the resolution of feature maps
• Upsample based on the nearest neighbor approach

U-Net

• Segmenting medical images
• Input Image -> Convolution -> RELU pooling
• Encoder - Similar to Image Classifier
• Upsampling through Decoder for same resolution output
• Upsampling - blobby feature map
• For every location distribution over classes
• Cross Entropy (Avg Over all Locations)

Keep Thinking!!!