motivation

Fitnets: Hints for thin deep nets

hint learning 利用网络中间层 teacher & student feature map L2 loss
分析 feature map 整体做 hint 不好，detectors care more about local regions that overlap with ground truth objects while classification models pay more attention to global context.

本文并不是对整个feature map 做的 hint learning，而是只对 gt附近的anchor 进行蒸馏，有一个 fine-grained feature limitation

红框：和 gt box IOU 最大
绿框：与 gt box IOU 大于下限阈值

method

对于每个 gt box，我们计算它与所有锚点之间的IOU，从而形成 W×H×K IOU map m
(W和H表示 feature map 的宽度和高度，K表示K个预设锚框)
然后我们找到最大的IOU值 M = max（m），乘以阈值因子ψ（超参数，default ψ=0.5），以获得滤波器阈值F =ψ∗M
利用 F 对 IOU map 进行过滤，OR操作保留大于F的位置得到 W×H mask
循环遍历所有 gt box 并得到 mask，得到 fine-grained imitation mask I

a full convolution adaptation layer
（after corresponding student model before calculating distance metric between student and teacher’s feature response）

teacher 和 student channel 数量不一定对齐
直接强迫 student 近似 teacher 的 feature 收益小

loss

we train student model to minimize the following objective

sss as student model’s guided feature map
ttt as corresponding teacher’s feature map
For each near object anchor location (i,j)(i, j)(i,j) on the feature map of width WWW and height HHH

Together with all estimated near anchor location(the imitation mask III), the distillation objective is to minimize:

I is the imitation mask
Np is the number of positive points in the mask
fadap(·) is the adaptation function.

Then the overall training loss of a student model is:

result

on faster rcnn

【Distill 系列：二】CVPR 2019 Distilling Object Detectors with Fine-grained Feature Imitation相关推荐

【没有哪个港口是永远的停留~ 论文简读】2019 Distilling Object Detectors with Fine-grained Feature Imitation
链接:2019 Distilling Object Detectors with Fine-grained Feature Imitation 代码:https://github.com/twangn ...
Distilling Object Detectors with Fine-grained Feature Imitation(2019 CVPR KD)
@[TOC](Distilling Object Detectors with Fine-grained Feature Imitation(2019 CVPR KD)) Abstract 我们表明, ...
【CVPR 2021】基于解耦特征的目标检测知识蒸馏：Distilling Object Detectors via Decoupled Features
[CVPR 2021]基于解耦特征的目标检测知识蒸馏:Distilling Object Detectors via Decoupled Features 论文地址: 主要问题: 主要思路: 具体实现 ...
[目标检测知识蒸馏2] [CVPR19] Distilling Object Detectors with Fine-grained Feature Imitation
[CVPR19] Distilling Object Detectors with Fine-grained Feature Imitation Introduction: 在 Faster RCNN ...
Distilling Object Detectors with Fine-grained Feature Imitation
Motivation 检测起更focus在物体出现的区域 Detectors care more about local near object regions. 物体出现的周围特征变化其实包含了更多 ...
Distilling Object Detectors via Decoupled Features
Abstract 相比于图像分类而言,目标检测器更加复杂,具有多个损失函数.而目前的的检测中,其主要将注意力集中在对象的区域中,但本文指出,从背景中提取的特征信息对于学生模型的学习也是必不可少的.且由 ...
Distilling Object Detectors with Fine-grained Feature Imitation论文解读
论文链接文章目录 1.简介 2.方法 2.1 模仿区域的估计 2.2 细粒度特征模仿 1.简介对于以往应用于知识蒸馏的目标检测存在两种方法.第一种来自于论文FitNets: Hints for T ...
论文笔记 Feature Selective Anchor-Free Module for Single-Shot Object Detection - CVPR 2019
2019 FSAF Feature Selective Anchor-Free Module for Single-Shot Object Detection Chenchen Zhu, Yihui ...
【CVPR 2019】Strong-Weak Distribution Alignment for Adaptive Object Detection
1 介绍此文章是[CVPR 2019]的文章,主要讲述了Strong-Weak的DA方法: 2 原文链接 <Strong-Weak Distribution Alignment for Ada ...

【Distill 系列：二】CVPR 2019 Distilling Object Detectors with Fine-grained Feature Imitation

motivation

method

loss

result

【Distill 系列：二】CVPR 2019 Distilling Object Detectors with Fine-grained Feature Imitation相关推荐

最新文章

热门文章