blue-oil/blueoil

Implement mixup data augmentation

Open

#430 创建于 2019年9月17日

在 GitHub 查看
 (0 评论) (1 反应) (0 负责人)Python (250 star) (85 fork)batch import
accuracy_improvementgood first issue

描述

  • For classification
  • For object detection

Mixup is an effective data augmentation, it can improve performance simply reported by many papers.

Paper: https://arxiv.org/pdf/1710.09412.pdf

References:

https://qiita.com/yu4u/items/70aa007346ec73b7ff05

https://arxiv.org/pdf/1902.04103.pdf

贡献者指南

技术栈
pythontensorflow
领域
machine learning
议题类型
feature
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
3
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
1-2 days
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
stale
清晰度议题是否清楚说明期望改动、验收标准和下一步。
mostly clear
前置要求
Python basicsDeep learning conceptsFamiliarity with data augmentation
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
40
研究方向
The issue requests implementing mixup data augmentation for classification and object detection tasks. The paper (arXiv:1710.09412) and reference implementations (e.g., Qiita article, CutMix paper) provide guidance. To begin, locate the existing data augmentation module in Blueoil (likely under blueoil/data augmentor.py or similar), and add a mixup function that blends images and labels. For object detection, adjust bounding boxes accordingly. Test the implementation against existing models.