Learn Coco Dateset with its annotations.

Overlook the dataset. We will count the categories for every pixels, and draw heatmap graphs for the counts.

• Download the annotations of Coco dataset;
• Put them into the folder of annotations;
• The overlook.py script will automatically count the histogram of objects' bbox in the annotations.
• The results will be saved into the folder of largeFiles.

Statistic the prior probabilities of categories, and posterior probabilities between each two categories.

• The prob_analysis.py script will compute the probability of categories, it will also compute the posterior probability between each two categories.

Use PCA and TSNE method to project the samples into regularized 3D space, the features are their categories vector in hot-encoding format.

• The category_space_analysis.py script will do the computation.

You will get images like these

• Count heat map of Occlude

  

• Histogram of categories

  

• Box graph of objects' area

  

• The base probabilities of categories

  

• The posterior probabilities of each two categories