[2Q6-GS-10] Vision, speech: Image analysis and application

[2Q6-GS-10-04] The effect of image resolution on inter-object relationship recognition

〇Ikuto Kurosawa¹, Tetsunori Kobayashi¹, Yoshihiko Hayashi¹ (1. Waseda University)

Keywords:inter-object relationship recognition , resolution of feature maps, scene graph

Recognition of the relationship between objects is a task of predicting the positional relationship or the action relationship that could hold between a pair of objects depicted in an image.
In this task, the visual feature of an object is required: feature maps obtained by CNN's convolutional layer and pooling layer are generally used.
However, they are considered insufficient to capture detailed information, such as the position of the human hand and the posture of the object, due to their low resolution, which is defined by the height and the width.
To address this issue, we propose to adopt Feature Pyramid Network (FPN) to increase the resolution of the feature map.
The experimental results proved that the use of the FPN was most effective in increasing the resolutions, contributing to improve the performance of the recognition of inter-object relationship recognition.