saliency-cropper

a toy saliency cropper based on gradient descent, implemented with pytorch. /inputs is the folder for input images, /outputs is the folder for output images.

Usage

python cropperlayer.py

Hyperparameters

ALPHA

• large ALPHA will encourage optimized larger bounding boxes
• default is 0.1.

BETA

• large BETA will encourage choosing proposal cluster with majority
• default is 0.5.

GAMMA

• large GAMMA will encourage choosing proposal with larger bounding boxes
• default is 10.0
• in this magnificat,we always choose maxim width bbox.

MAX_ITER

• controls the number of optimized iterations
• default is 100.

MINIMUM_THRESHOLD

• controls the filter of small bounding boxes
• default is 0.05.

KMEANS_K

• controls the number of proposal cluster
• we use it to filter outliers
• default is 5.