How to deal with the classes which have just a few samples in object detection?

[RangeKing]

How to deal with the classes which have just a few samples in object detection?
对于目标检测中样本很少的类别该怎么处理？

[RangeKing]

• hard example mining
• add more data
• loss function: focal loss
• For images, reduce sampling the data with many samples and resample the data with many samples.

More methods to search for long tail distribution

---

• 困难样本挖掘
• 增加数据
• 损失函数：focal loss
• 对图像样本来说，类别数据多的减少数据，类别数据少的重采样几遍

更多方法搜索长尾分布