Algorithm

Object searching

Because previous object-suggestion methods (e.g. selective search) tend not to be successful for detecting and counting flowers from a photograph of a meadow, which is my purpose to build this module, only traditional brute-force search with given sliding windows has been implemented. The bounding box size is determined by that of train data in the workflow below.

• X-means clustering of bounding box sizes in training data (initial number of cluster is 4)

• The center of each cluster is determined to be the size of the sliding window

Label estimation for each patches

After getting a patch with a sliding window, the semi-supervised object classifier runs to estimate its label. This is based on Gaussian Mixture Variational Auto Encoder (GMVAE) proposed by Rui Shu. You can see an MNIST experiment of GMVAE here.