Dense Teacher: Dense Pseudo-Labels for Semi-supervised Object Detection (2022-07-06T00:00:00.000000Z)

TL;DR

This work proposes replacing the sparse pseudo-boxes with the dense prediction as a united and straightforward form of pseudo-label as Dense Teacher, and introduces a region selection technique to highlight the key information while suppressing the noise carried by dense labels.

Abstract

To date, the most powerful semi-supervised object detectors (SS-OD) are based on pseudo-boxes, which need a sequence of post-processing with fine-tuned hyper-parameters. In this work, we propose replacing the sparse pseudo-boxes with the dense prediction as a united and straightforward form of pseudo-label. Compared to the pseudo-boxes, our Dense Pseudo-Label (DPL) does not involve any post-processing method, thus retaining richer information. We also introduce a region selection technique to highlight the key information while suppressing the noise carried by dense labels. We name our proposed SS-OD algorithm that leverages the DPL as Dense Teacher. On COCO and VOC, Dense Teacher shows superior performance under various settings compared with the pseudo-box-based methods.

Authors

Zeming Li

8 papers

Zheng Ge

5 papers

Jian‐Yuan Sun

3 papers

TL;DR

Abstract

Authors

References42 items

Distilling Object Detectors with Feature Richness

Semi-Supervised Object Detection with Adaptive Class-Rebalancing Self-Training

End-to-End Semi-Supervised Object Detection with Soft Teacher

Rethinking Pseudo Labels for Semi-Supervised Object Detection

Interactive Self-Training with Mean Teachers for Semi-supervised Object Detection

Humble Teachers Teach Better Students for Semi-Supervised Object Detection

Instant-Teaching: An End-to-End Semi-Supervised Object Detection Framework

Unbiased Teacher for Semi-Supervised Object Detection

DualBox: Generating BBox Pair with Strong Correspondence via Occlusion Pattern Clustering and Proposal Refinement

Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection

A Simple Semi-Supervised Learning Framework for Object Detection

PS-RCNN: Detecting Secondary Human Instances in a Crowd via Primary Object Suppression

A Simple Framework for Contrastive Learning of Visual Representations

FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence

Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adaptive Training Sample Selection

ReMixMatch: Semi-Supervised Learning with Distribution Alignment and Augmentation Anchoring

Self-Training With Noisy Student Improves ImageNet Classification

MixMatch: A Holistic Approach to Semi-Supervised Learning

FCOS: Fully Convolutional One-Stage Object Detection

CrowdHuman: A Benchmark for Detecting Human in a Crowd

Focal Loss for Dense Object Detection

Virtual Adversarial Training: A Regularization Method for Supervised and Semi-Supervised Learning

Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results

Feature Pyramid Networks for Object Detection

Temporal Ensembling for Semi-Supervised Learning

Regularization With Stochastic Transformations and Perturbations for Deep Semi-Supervised Learning

Deep Residual Learning for Image Recognition

Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks

Learning with Pseudo-Ensembles

ImageNet Large Scale Visual Recognition Challenge

Microsoft COCO: Common Objects in Context

IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2022, New Orleans, LA, USA, June 19-20, 2022

Consistency-based Semi-supervised Learning for Object detection

Dropout: a simple way to prevent neural networks from overfitting

Pseudo-Label : The Simple and Efficient Semi-Supervised Learning Method for Deep Neural Networks

The PASCAL visual object classes challenge 2006 (VOC2006) results

The Pascal Visual Object Classes Challenge 2006 ( VOC 2006 ) Results

Since the learning region is selected, unsupervised learning for regression branch can be easily achieved

By modifying the learning region, we can easily achieve Hard Negative Mining by selecting extra samples (see Fig. 4)

We propose a united form of pseudo-label named DPL to better fit the semi-supervised setting and the Dense Teacher framework to apply the DPL on one-stage detector

The student model then takes the labeled data for vanilla training and calculates supervised loss L s , while the unlabeled data together with pseudo-labels are used to produce unsupervised loss L u

“Anchor” stands for “anchor point” in anchor-free detectors and “anchor box” in anchor-based detectors

Field of Study

Journal Information

Name

Volume

Venue Information

Name

Type

URL

Alternate Names