COCO Challenges & Dataset - Stuff Segmentation

[junxnone]

Stuff Segmentation - Semantic Segmentation [2017 ~ Now]

Reference

• [ ] COCO 2019 Stuff Segmentation Task
• [ ] Github - The COCO-Stuff dataset
• [ ] Github - nightrome/cocostuff10k
• [ ] Github - COCO Stuff API
• [ ] paper - 2016 - COCO-Stuff: Thing and Stuff Classes in Context
• [ ] pngToCocoResultDemo.py
• [ ] RLE format
• [ ] metrics paper - Fully Convolutional Networks for Semantic Segmentation

Brief

Classes

• 91 stuff classes
• other class

Annotation Json Format

[{
"image_id": int, 
"category_id": int, 
"segmentation": RLE, 
"score": float,
}]

RLE Format - Run-length encoding

• MASK API 自定义的运行长度编码
  • 紧凑地存储掩码
  • 高效地执行掩码计算
  • 阵列 - O(n)
  • RLE - O(n^0.5)

Metrics

• IoU=TP/(TP+FP+FN)
• mIoU 指每类 IoU 的均值

defined from paper Fully Convolutional Networks for Semantic Segmentation

[junxnone]

244 #686