UX-Decoder / DINOv

[CVPR 2024] Official implementation of the paper "Visual In-context Learning"
312 stars 10 forks source link

为什么在训练过程中需要positive and negative visual prompt samples? #27

Open dbcSep03 opened 1 week ago

dbcSep03 commented 1 week ago

如题,有点不太明白

FengLi-ust commented 5 days ago

没有negative的话,模型只能检测正样本。比如图片里面没有狗,你输入狗的prompt,模型还是会预测出一个错误的mask,因为训练中所有prompt都会是positive,模型一定要输出mask才行。