A semi-supervised object segmentation method by using adaptive memory network
REM extracts the features of objects by focusing on the object area and filtering out background information.
cd /extensions//reg_att_map_generator
python setup.py install --user
from extensions.reg_att_map_generator import RegionalAttentionMapGenerator
self.ram = RegionalAttentionMapGenerator()
att_map, bbox = self.ram(curr_mask, n_object)
The features of all objects are encapsulated in a single feature map containing the salient information, significantly reducing the number of features that need to be stored.
ID enables the integration of multiple-object features within a unified feature space.
FCM compresses the outdated features into the feature containing the salient information.
LMM enhances segmentation speed and accuracy through local matching between the query feature of the current frame and the key features of the memoried historical frames.