(Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

(Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

Ref

1. 提出研究的问题

对于few-shot分类而言,可以采用学习prototype的方式学习一个中心,但是如果用于目标检测,很难匹配和定位一个物体。因为在图像中会有很多的干扰因素和非目标物体。

2. 提出方案

  • Feature Extractor module

    Feature Extractor从 query images中提取meta features ,表示为FRwhmF \in R^{w*h*m}

    文章中的的Extractor是采用YOLO v2的darknet, 用D(I)D(I)表示;

  • Reweighting Module d从support set中提取全局信息并embeded成reweighting coefficient,表示为ω(i)Rm\omega (i) \in R^m

    文章中reweighting是简单的几层网络层,用M(Ii,Mi)M(I_i,M_i)表示,下图右下角IiI_i表示图像,MiM_i表示mask;

  • Prediction Module

    跟YOLO v2相似,预测的值增加one-versus-all classification score cic_i表示对应目标属于类别i的概率;

    总共预测六个值:{oi,xi,yi,hi,wi,ci}=p(Fi)i=1,...,N\{o_i,x_i,y_i,h_i,w_i,c_i\} = p(F_i) i= 1,...,N ;

(Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

  • 为什么有检测新目标的能力? https://github.com/bingykang/Fewshot_Detection

    根据paper的损失函数:

    (Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

    (Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

    (Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

    我认为base training过程主要是训练base -> novel的能力和localization的能力;

    而在few-shot fine-tuning更像是pretrained模型通过support sets预训练fine-tune的过程,当然两者的数据集不能有交集。

3. 总结

模型其实是做一种Transfer Learning的思想,与原来detection不同的是,训练方式不同,数据组成的方式不同。

数据分为base classes和novel classes,对于每一个task或者episode,包含Support sets和Query sets两部分,

(Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

网络模型和损失函数上述2.提出方案所示;

整个学习过程分为两个阶段:

  • base training phase

    在此阶段,仍采用base classes进行episode的方式训练,目的是让模型学会通过reweighting vector找到感兴趣的区域;

  • few-shot fine-tuning

    在此阶段,训练数据包括base classes和novel classes,对于novel classes来说,只有k个标注的bounding boxes,对于base classes的数据,也有k个bounding boxes,训练过程和第一阶段相同。support images和boundingbox的数据从base classes和novel classes的数据中随机选取,进行fine-tune;

4. 实验结果

(Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

(Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

  • 分析结果

    (Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

    表明作者提出的few-shot模型在速度和性能上都能超过原来的YOLO v2模型;

  • Which layer output features to reweight

    (Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

  • Loss functions

    (Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

  • Input form of reweighting module

    (Few-shot Detection)Review: Fews-shot Object Detection Feature Reweighting

这么多的对比实验,实践出真知啊。