PROJECT TITLE :

Focal Boundary Guided Salient Object Detection

ABSTRACT:

The use of deep convolutional networks has significantly improved the performance of salient object segmentation. However, these networks often yield blob-like maps of saliency that do not accurately represent the boundaries of individual objects. After several pooling procedures, their feature maps have a restricted spatial resolution, which may impede downstream applications that demand precise object forms. A unique deep model-Focal Boundary Guided (Focal-BG) network is proposed to overcome this problem. Salient object masks are segmented using our model, while salient object borders are detected using our model's algorithm. We believe that extra information regarding the borders of an object can aid in the exact identification of the object's shape. A refinement process for mask prediction and the usage of the focus loss in the learning of hard boundary pixels are also included in our model. We run a wide range of experiments to test our model. This network routinely surpasses the state of the art approaches in five important benchmarks. In our extensive research, we show that our joint modelling of salient object boundaries and masks helps to better capture the form features, particularly in the vicinity of object boundaries.