CVPR 2021 Object Detection

一. 关于3D有26篇:

3DIoUMatch: Leveraging IoU Prediction for Semi-Supervised 3D Object Detection
ST3D: Self-Training for Unsupervised Domain Adaptation on 3D Object Detection
SRDAN: Scale-Aware and Range-Aware Domain Adaptation Network for Cross-Dataset 3D Object Detection
Unsupervised Object Detection With LIDAR Clues:利用雷达点云做无监督目标检测

二. 关于2D有55篇:

有几篇没有review,例如音频等考虑在内的一些论文直接略过了。

Semi-Supervise(6篇):

  1. Instant-Teaching: An End-to-End Semi-Supervised Object Detection Framework:从数据增强与训练策略的角度出发,完善teacher-student模式;
  2. Points As Queries: Weakly Semi-Supervised Object Detection by Points:数据标注采用了部分标bbox、部分在object上标点和类别的方式,模型采用了改进的DETR
  3. Humble Teachers Teach Better Students for Semi-Supervised Object Detection:EMA相关;
  4. Data-Uncertainty Guided Multi-Phase Learning for Semi-Supervised Object Detection:数据分析相关;
  5. Interactive Self-Training With Mean Teachers for Semi-Supervised Object Detection:对检测任务的训练过程中出现的,对同一张图片产生不同预测bbox之间存在的冲突进行处理(未完);
  6. Interpolation-Based_Semi-Supervised_Learning_for_Object_Detection:数据分析相关,bbox之间进行插值方式的融合;

positive unlabeled learning:
Positive-Unlabeled Data Purification in the Wild for Object Detection:从大量wild未标注数据中提纯数据,加入到训练中,提点(未完)

Cross Domain & Domain adaptive(4篇)

  1. Unbiased Mean Teacher for Cross-domain Object Detection:teacher-student模式,关于EMA,PASCAL VOC作为source domain(自然数据集),target domain是卡通风格数据集;
  2. Informative and Consistent Correspondence Mining for Cross-Domain Weakly Supervised Object Detection:PASCAL VOC作为source domain(自然数据集),target domain是卡通风格数据集,model沿用faster-RCNN;
  3. MeGA-CDA: Memory Guided Attention for Category-Aware Unsupervised Domain Adaptive Object Detection:model沿用faster-RCNN,VGG16 backbone,SGD优化器,数据集都是真实场景下的,一个是无人驾驶场景,一个是城市景观数据集,互为S、D;
  4. Domain-Specific Suppression for Adaptive Object Detection:从训练权重与梯度的角度出发讨论,网络为ResUnit相关,数据集中涉及域之间的天气差异(有没有雾)、相机设置差异、真实图像与合成图像间差异

关于Human Object Interaction(5篇)
有三篇都是基于transformer做的

  1. Affordance Transfer Learning for Human-Object Interaction Detection:检测时采用faster-RCNN结构
  2. Glance and Gaze: Inferring Action-aware Points for One-Stage Human-Object Interaction Detection:提出了一个 GGNet

显著性目标检测(3篇)

  1. Group Collaborative Learning for Co-Salient Object Detection:联合显著性目标检测
  2. Calibrated RGB-D Salient Object Detection:将图像RGB stream与深度stream分别编码
  3. Weakly Supervised Video Salient Object Detection:双向的ConvLSTM,弱监督,图像标注两笔,一笔画在显著性目标上一笔在背景上
  4. Uncertainty-aware Joint Salient Object and Camouflaged Object Detection

少样本学习(9篇)
例:Dense Relation Distillation With Context-Aware Aggregation for Few-Shot Object Detection:少样本目标检测,框架基于元学习,处理检测中物体间遮挡、外观变化等问题;
Hallucination Improves Few-Shot Object Detection:网络结构有基于Faster-RCNN

关于Transformer(7篇)

  1. Adaptive Image Transformer for One-Shot Object Detection:使用patch作为query(未完);
  2. UP-DETR: Unsupervised Pre-Training for Object Detection With Transformers:使用patch作为query,从原始图像中查询这些random裁剪的patch(未完);
  3. HOTR: End-to-End Human-Object Interaction Detection With Transformers:从图像中直接推断三元组,例如:<人,凳子,坐>;
  4. End-to-End Human Object Interaction Detection With HOI Transformer:也是关于人物交互的
  5. QPIC: Query-Based Pairwise Human-Object Interaction Detection With Image-Wide Contextual Information:也是关于人物交互的
  6. Open-Vocabulary Object Detection Using Captions:建立一个图像和文字标签的空间,建立空间的时候使用了transformer,使用边界框注释对有限的对象类别集进行训练
  7. Points As Queries: Weakly Semi-Supervised Object Detection by Points:改进的DETR

搜索(4篇)

  1. GAIA: A Transfer Learning System of Object Detection That Fits Your Needs:自动搜索合适的目标检测网络结构;
  2. Scale-Aware Automatic Augmentation for Object Detection:自动搜索合适的数据增强方式,关注scale的aug,从image-level与bbox-level去aug数据,例如image-level采用zoom-in zoom-out;
  3. OPANAS: One-Shot Path Aggregation Network Architecture Search for Object Detection:自动搜索网络结构用于目标检测(NARS)
  4. MobileDets: Searching for Object Detection Architectures for Mobile Accelerators:搜索移动端的目标检测网络结构

长尾问题(2篇)

  1. Equalization Loss v2: A New Gradient Balance Approach for Long-Tailed Object Detection:提出长尾目标检测问题在于正负梯度不平衡,提出了EQL_v2 loss用以平衡每个类别训练过程,模型采用了maskrcnn与Cascade
  2. Adaptive Class Suppression Loss for Long-Tail Object Detection:打破手工分组的局限性,设计了一种新的自适应类抑制损失(ACSL)算法,模型为Faster R-CNN with ResNet50-FPN backbone

关于生成对抗(2篇)

  1. Class-Aware Robust Adversarial Training for Object Detection:加入noisy扰动干扰检测
  2. Robust and Accurate Object Detection via Adversarial Learning:使用对抗方法robust检测器

另 一些各种各样角度的论文:

航空图像目标检测:
ReDet: A Rotation-equivariant Detector for Aerial Object Detection:关注旋转不变性特征

关于量化:
AQD: Towards Accurate Quantized Object Detection:保证精度的条件下在所有网络结构层使用纯整数进行推理,已应用于FCOS、RetinaNet

关于Attention
Dynamic Head: Unifying Object Detection Heads with Attentions:对目标检测任务检测头的注意力进行可视化(未完)

关于知识蒸馏:
General Instance Distillation for Object Detection

主动学习:
Multiple Instance Active Learning for Object Detection

Mutual Graph Learning for Camouflaged Object Detection:MGL将一幅图像解耦为两个特定于任务的特征映射——一个用于大致定位目标,另一个用于准确捕捉其边界细节——并通过图来反复推理它们的高阶关系,模型ResNet-FCN

Improved Handling of Motion Blur in Online Object Detection:关注运动模糊问题,模型结构沿用resnet50 FPN,faster-RCNN

Neural Auto-Exposure for High-Dynamic Range Object Detection:关于曝光图像等的目标检测(未完)

End-to-End Object Detection With Fully Convolutional Network:抛弃NMS,采用3D最大卷积,提出Prediction-aware One-To-One (POTO) label assignment(未完)

OTA: Optimal Transport Assignment for Object Detection(暂时没看懂)

Depth from Camera Motion and Object Detection:使用faster-RCNN为检测器,关于相机运动情况下的物体深度估计问题

Beyond Bounding-Box: Convex-hull Feature Adaptation for Oriented and Densely Packed Object Detection:可形变卷积,anchor free,利用点集形式而非bbox表示一个物体的轮廓(未完)

Towards Open World Object Detection:基于聚类和能量的模型

UniT: Unified Knowledge Transfer for Any-shot Object Detection and Segmentation:基于Faster-RCNN & Mask-RCNN,弱监督

Generalized Focal Loss V2: Learning Reliable Localization Quality Estimation for Dense Object Detection:focal loss的改进版,处理密集场景下的目标检测(未完)

Sparse R-CNN: End-to-End Object Detection with Learnable Proposals:提出了Sparse R-CNN(未完,这论文要细看)

你可能感兴趣的:(深度学习,深度学习)