2021CVPR | 2D目标检测

2021 CVPR 2D目标检测

1. You Only Look One-level Feature

paper: https://arxiv.org/abs/2103.09460

code: https://github.com/megvii-model/YOLOF

描述：在单阶段目标检测方面一次突破性的创新，它针对单阶段目标检测中的FPN(特征金字塔)进行了深入的分析并得出：FPN最重要的成分是分而治之的处理思路缓解了优化难问题。针对FPN的多尺度特征、分而治之思想分别提出了Dilated编码器提升特征感受野，Uniform Matching进行不同尺度目标框的匹配；结合所提两种方案得到了本文的YOLOF，在COCO数据集上，所提方案取得了与RetinaNet相当的性能且推理速度快2.5倍；所提方法取得了与YOLOv4相当的性能且推理速度快13%。

2. General Instance Distillation for Object Detection

paper:https://arxiv.org/abs/2103.02340

描述：GID提出了一种基于检测任务的新型蒸馏方法。通过从teacher和studnet中分别提取general instance (GI)，并提出GISM模块自适应选择差异大的instance进行feature-based、relation-based以及response-based蒸馏。本方法首次将关系型知识蒸馏应用于检测框架，且将蒸馏目标从独立考虑的正负样本蒸馏统一为更本质GI蒸馏，过程中不依赖于GT，且达到SOTA。

3. Points as Queries: Weakly Semi-supervised Object Detection by Points

描述：在不增加标注成本的条件下，提升检测器的性能，是本文研究的目标。本文选择少量边界框辅以大量点标注的方式训练检测器。选择点标注是因其信息丰富：包含实例的位置和类别信息，同时标注成本低。本文通过将点编码器扩展至DETR的方式，提出Point DETR，整体框架为：通过边界框数据训练Point DETR；将点标注编码为查询，预测伪框；通过边界框和伪框数据，训练学生模型。在COCO数据集上，仅使用20%完全标注的数据，我们的检测器可达33.3AP，超过基线2.0AP。

4. Probabilistic two-stage detection

paper：https://arxiv.org/abs/2103.07461

code：https://github.com/xingyizhou/CenterNet2

5. ReXNet: Diminishing Representational Bottleneck on Convolutional Neural Network

paper: https://arxiv.org/abs/2007.00992

code: https://github.com/clovaai/rexnet

6. UP-DETR: Unsupervised Pre-training for Object Detection with Transformers

paper：https://arxiv.org/abs/2011.09094

code：https://github.com/dddzg/up-detr

7. Coordinate Attention for Efficient Mobile Network Design

paper：https://arxiv.org/abs/2103.02907

code（刚刚开源）：https://github.com/Andrew-Qibin/CoordAttention

8. Generalized Focal Loss V2: Learning Reliable Localization Quality Estimation for Dense Object Detection

paper：https://arxiv.org/abs/2011.12885

code：https://github.com/implus/GFocalV2

9. Dense Label Encoding for Boundary Discontinuity Free Rotation Detection

paper：https://arxiv.org/abs/2011.09670

code：https://github.com/yangxue0827/RotationDetection

10. Inception Convolution with Efficient Dilation Search

paper：https://arxiv.org/abs/2012.13587

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

Paper: https://arxiv.org/abs/2103.11402

Code: None

12. OPANAS: One-Shot Path Aggregation Network Architecture Search for Object Detection

Paper: https://arxiv.org/abs/2103.04507

Code: https://github.com/VDIGPKU/OPANAS

13. Semantic Relation Reasoning for Shot-Stable Few-Shot Object Detection

Paper: https://arxiv.org/abs/2103.01903

Code: None

14. There is More than Meets the Eye: Self-Supervised Multi-Object Detection and Tracking with Sound by Distilling Multimodal Knowledge

Homepage: http://rl.uni-freiburg.de/research/multimodal-distill

Paper: https://arxiv.org/abs/2103.01353

Code: http://rl.uni-freiburg.de/research/multimodal-distill

15. Multiple Instance Active Learning for Object Detection

Paper（链接失效）: https://github.com/yuantn/MIAL/raw/master/paper.pdf

Code: https://github.com/yuantn/MIAL

16. Towards Open World Object Detection

Paper: https://arxiv.org/abs/2103.02603

Code: https://github.com/JosephKJ/OWOD