InstaBoost

A python implementation for paper: "InstaBoost: Boosting Instance Segmentation Via Probability Map Guided Copy-Pasting". The code has been released on PyPI for easier installation and implementation.

Install

pip install instaboost

Only tested on python3 and Linux. Windows is not supported since pycocotools do not support Windows.

Get Started

There are two main interface in InstaBoost.

InstaBoostConfig

cfg = InstaBoostConfig(action_candidate: tuple,
                       action_prob: tuple, 
                       scale: tuple, 
                       dx: float, 
                       dy: float,
                       theta: tuple, 
                       color_prob: float, 
                       heatmap_flag: bool)

Parameters:

action_candidate: tuple of action candidates. 'normal', 'horizontal', 'vertical', 'skip' are supported

action_prob: tuple of corresponding action probabilities. Should be the same length as action_candidate

scale: tuple of (min scale, max scale)

dx: the maximum x-axis shift will be (instance width) / dx

dy: the maximum y-axis shift will be (instance height) / dy

theta: tuple of (min rotation degree, max rotation degree)

color_prob: the probability of images for color augmentation

heatmap_flag: whether to use heatmap guided

Output:

cfg: a InstaBoostConfig instance

get_new_data

new_ann, new_img = get_new_data(ori_anns: list, 
                                ori_img: np.ndarray, 
                                config: InstaBoostConfig, 
                                background: np.ndarray)

Parameters:

ori_anns: list of coco-style annotation dicts

ori_img: image corresponding to ori_anns

config: a InstaBoostConfig instance. If None, the default parameters will be used

background: if not None, this background image will be used for augmentation

Output:

new_ann: ori_anns after augmentation without changes in length of list or keys of dicts

new_img: ori_img after augmentation without changes in shape

Samples and models

We show how to implement our method on two main segmentation frameworks: Detectron and mmdetection in repo InstaBoost. Results and models trained with InstaBoost are available in the Model zoo.