Main

用一组数量少且可学习的proposal（100个）代替来自RPN模块大量的proposal。

Method

Input

one image, proposal box (learnable), proposal features (learnable)

Backbone

FPN based on ResNet (ResNet50), Channel = 256

Learnable proposal box

Shape: [N, 4]; N means box numbers (100); 4 means 4-d parameters, including normalized center coordinates (ranging from 0 to 1), height and width

Learnable proposal feature

Notes that boxes only provide localization and some information details are lost, such as object pose and shape.
Shape: [N, d]; d means high dimension (256).
box.shape = proposal feature.shape

Dynamic instance interactive head

Before: 通过RoIAlign得到RoI feature；proposal box、proposal feature和RoI feature的关系是一一对应的。
Structure:
两个1*1卷积（ReLU激活函数）：使用proposal feature，生成两个卷积的参数。使用RoI feature和两个卷积的参数，生成object feature。
Iteration structure: The newly generated object boxes and object features will serve as the proposal boxes and proposal features of next stage.

Loss

L_cls是focal loss，L_L1和L_giou分别为L1loss和GIOUloss。