Fast Online Object Tracking and Segmentation: A Unifying Approach

Abstract

perform both visual object tracking and semi-supervised video object segmentation.

Once trained solely relies on single bounding box initialization and operates online, producing class-agnostic object segmentation masks. (35 fps, SOTA on VOT 2018)

Introduction

Three tasks such as Mask, Box and Score.

One task is to learn a measure of similarity between the target object and multiple candidates in the window. To refine information learn two further tasks: bounding box regression using RPN and class-agnostic binary segmentation.

Fully Convolutional Siamese framework
RoW: response map as a response of a candidate window

References

[ ] A novel performance evaluation methodology for single-target trackers. (IEEE 2016)
[ ] Eco: Efficient convolution operators for tracking. (IEEE 2017)
[ ] A benchmark dataset and evaluation methodology for video object segmentation. (IEEE 2017)
[ ] Fully-convolutional siamese networks for object tracking. (ECCV 2016)
[ ] High performance visual tracking with siamese region proposal network. (IEEE 2018)

jiunbae / Lab