Constrained fixation point based segmentation via deep neural network

Personal Fixations-Based Object Segmentation with Object Localization and Boundary Preservation

“…Khosravan et al [32] integrated fixations into the medical image segmentation and proposed a Gaze2Segment system. Li et al [25] constructed a dataset where all fixations fall in objects (i.e. constrained fixations), and proposed a CNNbased model to simulate the human visual system to segment objects based on fixations.…”

“…These studies have promoted the development of fixationsbased object segmentation. However, all the fixations in [22], [25], [30], [31] fall in objects, which are hardly guaranteed in practice. These methods [22], [25], [30], [31] will get stuck in the ambiguity of unconstrained fixations, especially of personal fixations.…”

“…However, all the fixations in [22], [25], [30], [31] fall in objects, which are hardly guaranteed in practice. These methods [22], [25], [30], [31] will get stuck in the ambiguity of unconstrained fixations, especially of personal fixations. For [23], [24], they are based on region proposal and cannot obtain accurate results.…”

“…The first category is that all fixations fall in the objects/foreground, i.e. the constrained fixation map in [25]. The second category is that some fixations fall in the background, namely the unconstrained fixation map.…”

“…The typical manners of interaction, such as scribbles [7]- [11], clicks [12]- [17] and bounding boxes [18]- [21] for interactive image segmentation, are explicit behaviors without interference. By contrast, fixations are implicit [22]- [25], and their convenience comes with interaction ambiguity. Concretely, the positive and negative labels of scribbles and clicks are deterministic.…”