Human-like Controllable Image Captioning with Verb-specific Semantic Roles(具有动词语义角色的类人可控图像字幕生成)

 前人的缺陷：

CIC works mainly focus on (1)subjective control signals,(2)objective control signals  or (1) Content-controlled (2) Structure controlled。

almost all existing objective control signals have overlooked two indispensable characteristics of an ideal control signal:

  1) Event-compatible:all visual contents referred to in a single sentence should be compatible with the describe activity.

  2) Sample-suitable: the control signals should be suitable for a specific image sample.

论文的创新点：

propose a new event-oriented objective control signal, Verb-specific Semantic Roles (VSR), to meet both event-compatible and sample-suitable requirements simultaneously。

VSR consists of a verb and some user-interested semantic roles。

Grounded Semantic Role Labeling: visual features of all grounded proposal sets。

Semantic Structure Planner: hierarchical semantic structure learning model, which aims to learn a reasonable sequence of sub-roles S。

Verb-specific Semantic Roles = Grounded Semantic Role Labeling  υ  Semantic Structure Planner

 

 step：we first use GSRL and SSP to obtain semantic structures and grounded regions features: (Sa; Ra) and (Sb; Rb).

Then,as shown in Figure above, we merge them by two steps。

  (a) find the sub-roles in both Sa and Sb which refer to the same visual regions 

  (b) insert all other sub-roles between the nearest two selected sub-roles

模型架构:

Faster R-CNN(ResNet-101) + Controllable LSTM + Controllable UpDn + SCT

原文: https://arxiv.org/abs/2103.12204