DTG-SSOD: The latest semi-supervised detection framework, Dense Teacher (with paper download)

关注并星标

从此不迷路

计算机视觉研究院

公众号ID｜ComputerVisionGzq

学习群｜扫码在主页获取加入方式

论文地址：https://arxiv.org/pdf/2207.05536.pdf

计算机视觉研究院专栏

作者：Edison_G

“从Sparse to dense”The paradigm makesSSODprocess is complicated,While ignoring the powerful direct、Intensive teacher supervision

01

概述

Mean-Teacher (MT) Scheme in semi-supervised object detection (SSOD) 中被广泛采用.在MT中,Final predictions by teachers（例如,In non-maximally inhibited (NMS) after post-processing）The provided sparse pseudo-labels intensively supervise students through hand-crafted label assignments.然而,“从Sparse to dense”The paradigm makesSSODprocess is complicated,While ignoring the powerful direct、Intensive teacher supervision.

在今天分享中,The researchers attempted to supervise the training of students directly using intensive instruction from teachers,即“dense to dense”范式.具体来说,The researchers proposed the inverseNMS聚类(INC)and rank match(RM)to instantiate dense supervision,without the widely used traditional sparse pseudo-labels.INCGuide students inNMSgroup the candidate boxes into clusters like a teacher,This is done by learning from the teacherNMSThe grouping information displayed in the program is realized.在通过INCAfter getting the same grouping scheme as the teacher,学生通过Rank MatchingFurther mimicking the teacher's ranking distribution among clustered candidates.

通过提出的INC和RM,将Dense Teacher GuidanceIntegrated into semi-supervised object detection（称为“DTG-SSOD”）中,Sparse pseudo-labels are successfully discarded,And achieves more informative learning on unlabeled data.在COCO基准测试中,新方法的DTG-SSODState-of-the-art performance is achieved at various labeling ratios.例如,在10%at the labeling rate,DTG-SSODWill monitor the baseline from26.9提高到35.9mAP,比之前的最佳方法Soft Teacher高19个百分点.

02

新框架

Comparison of teacher-supervised signals：下图(a)The previous method was performed on teachersNMSand fractional filtering to obtain sparse pseudo-labels,This further translates to intensive supervision of students through label assignment;下图(b)提出的DTG-SSODDirectly adopt the intensive predictions of teachers as intensive guidance for students.

Sparse-to-dense Paradigm

Task Formulation

SSOD的框架如下图(a)所示.Mean-TeacherScenarios are common practice with previous technologies,实现了端到端的训练,Passed after each training iterationEMABuild teachers from students.Teachers will be weakly enhanced（For example flip and resize）image as input to generate pseudo labels,Students apply strong reinforcement（例如剪切、几何变换）进行训练.Robust and appropriate data augmentation plays an important role,It not only increases the difficulty of students' tasks and alleviates the problem of overconfidence,It also enables students to remain invariant to various input perturbations,This enables robust representation learning.

Sparse-to-dense Baseline

所有以前的SSODThe methods are all based on a sparse-to-dense mechanism,which generates sparse pseudo-boxes with class labels,to serve as the ground truth for student training.It comes with a confidence based threshold,Among them only keep with high confidence（例如,大于0.9）的伪标签.This makes foreground supervision on unlabeled data much sparser than on labeled data,因此,The class imbalance problem is thereSSOD中被放大,It seriously hinders the training of the detector.

为了缓解这个问题,The researchers took advantage of some of the strengths of previous work：Soft Teacherthe mixing ratior设置为1/4,in order to sample more unlabeled data in each training batch,This brings the number of foreground samples on unlabeled data close to labeled data;Unbiased Teacher用Focal lossInstead of cross-entropy loss,Thereby reducing the gradient contribution of simple examples.

These two improvements,That is, the appropriate mixing ratior(1/4)和Focal loss,Both are used for the sparse-to-dense baseline and the researchers' dense-to-denseDTG 方法.Because the teacher only provides sparse pseudo-labels,This further translates into intensive supervision of student training,这些方法被称为“Sparse to dense”范式.理论上,新提出的SSODThe method is independent of the detection framework,Can be applied to single-stage and two-stage detectors.For a fair comparison with previous works,使用Faster RCNNas the default detection framework.

03

实验

displayed as a table,Under the full tag data setting,新提出的DTG-SSOD大大超过了以前的方法,beyond at least1.2mAP.Follow the previous practice,The researchers also applied weak boosting to the labeled data,并获得了40.9mAPstrong supervised baseline.Even based on such a strong baseline,DTG-SSOD仍然获得了+4.8mAP的最大改进,达到了45.7mAP,This verifies the effectiveness of the new method when the amount of labeled data is large.

研究者在30kA checkpoint is used for analysis at the iteration.Student training labels provided by sparse pseudo-labels are carefully compared with researcher-intensive teacher guidance.(a)sparse-to denseParadigms and researchersdense-to-denseThe paradigm brings different training labels to student samples.(b)Teachers assign higher marks to high-quality candidates,This preserves the exact box.

Some visual examples to demonstrate the advantages of the newly proposed method over the traditional sparse-to-dense paradigm.(a-b)For the same student proposal,The new dense-to-dense paradigm and the traditional sparse-to-dense paradigm will assign different labels.很明显,The new dense-to-dense paradigm can assign more precise and reasonable training labels.(c)Teachers are better at modeling the relationships of cluster candidates than students.

The summary of transformations used in weak and strong augmentation

Today is Army Day,Use an appropriate onedemoEnd today's lecture.

 THE END 

转载请联系本公众号获得授权

计算机视觉研究院学习群等你加入！

我们开创“计算机视觉协会”知识星球两年有余,也得到很多同学的认可,最近我们又开启了知识星球的运营.我们定时会推送实践型内容与大家分享,在星球里的同学可以随时提问,随时提需求,我们都会及时给予回复及给出对应的答复.

ABOUT

计算机视觉研究院

计算机视觉研究院主要涉及深度学习领域,主要致力于人脸检测、人脸识别,多目标检测、目标跟踪、图像分割等研究方向.研究院接下来会不断分享最新的论文算法新框架,我们这次改革不同点就是,我们要着重”研究“.之后我们会针对相应领域分享实践过程,让大家真正体会摆脱理论的真实场景,培养爱动手编程爱动脑思考的习惯！

VX：2311123606

往期推荐 

• Yolov7：最新最快的实时检测框架,最详细分析解释（附源代码）

• Yolov7实战,Real-time target detection on the web page

• 打假Yolov7的精度,Not all papers are authentic
  

• ECCV 2022丨Lightweight model architecture is on fire,力压苹果MobileViT（Code and paper downloads are attached）
  

• 最新的目标检测的深度架构 | 参数少一半、速度快3倍+（文末福利）
  

• SSD7 | 对嵌入式友好的目标检测网络,产品落地

• 精度提升方法：自适应Tokens的高效视觉Transformer框架（已开源）
  

• ONNX 浅析：如何加速深度学习算法工程化？

• 劲爆！YOLOv6又快又准的目标检测框架开源啦（附源代码下载）

• FastestDet：比yolov5更快！更强！全新设计的超实时Anchor-free目标检测算法（附源代码下载）
  

• 目前精度最高效率最快存储最小的目标检测模型（附源码下载）

• CVPR小目标检测：上下文和注意力机制提升小目标检测（附论文下载）