In general, if images are captured by different equipment or in different settings: you could use histogram equalization, which would adjust the brightness of all images to the same level, or you could use a normalization technique such as batch normalization or feature standardization, which would adjust the color and contrast of all images.
Image Localization
Image Localization:
Localization is the task of finding a single object in an image.
Image Classification+Localization:
Localization can be combined with classification to, not only find the location of an object but, also, to classify it into one of different classes.
Structure:
Input: Image
Output: A vector of 4 coordinates of the bounding box.
Applications:
Smart Cropping
Regular Object Extraction (as a pre-processing step)
Human Pose Estimation: Represent pose as a set of 14 joint positions
Approaches
Localization as a Regression Problem:
Since we are concerned with returning real-valued numbers (the bounding box coordinates), we use a method that is suitable for the task, Regression.
Algorithm:
Use any classification architecture
Attach two Fully Connected Layers, one for Classification and one for Localization
Backpropagate through the whole network using cross-entropy loss and L2 loss respectively.
