3rd Place Solution for Open Images 2019 - Visual Relationship
Authors: Roman Solovyev; Weimin Wang; Vladislav Golubev; Arthur Stsepanenka; Nikolay Sergievskiy
Abstract: In this technical report, we discuss our 3rd place solution for the Visual Relationship track of Open Images 2019 competition. Specially, we demonstrate the effectiveness of using two models - ‘Relationship Model‘ and ‘Attribute Model‘ - to solve the task. In particular, for the ‘Relationship Model‘, we have shown in our results that leveraging on the high quality bounding boxes of individual objects, a gradient boost based model can achieve very good performance of predicting pairs of relations between objects based on carefully designed geometric features extracted from those individual boxes.