Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection

本文揭示了长尾目标检测中的主要问题是正负之间的梯度不平衡,并提出均衡损失(equalization loss)v2 表现SOTA!性能优于BAGS、Forest R-CNN等网络,代码即将开源!

注:文末附【目标检测】学习交流群

Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection

  • 作者单位:同济大学, 商汤科技, 清华大学
  • 论文:https://arxiv.org/abs/2012.08548
  • 代码:https://github.com/tztztztztz/eqlv2

Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection_第1张图片

Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection_第2张图片

Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection_第3张图片
Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection_第4张图片
Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection_第5张图片
Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection_第6张图片

目标检测交流群

CVer已建立目标检测微信交流群,已经上千人!想要进目标检测学习群的同学,可以直接加微信号:CVer5555。加的时候备注一下:目标检测+学校+昵称,即可。然后就可以拉你进群了。

Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection_第7张图片

你可能感兴趣的:(计算机视觉论文速递,目标检测,开源项目,深度学习,人工智能,机器学习,计算机视觉,自动驾驶)