Tan SU
Tan SU
I'm doing fp16 training recently and maybe I can fix the overwritten issue (introduced by #1013 ) in a more elegant way. No need to merge yet. @HAOCHENYE _Update: new...
Any ideas about the CI failure?
I moved the unrelated changes to other PRs and wrote a new unitest for this use case.
这一部分是用来定量评估拼接质量的，当时没有整理，时间也很久了，这部分当时也不是我写的，所以请参考论文吧，主要思想还是很简单的。大致是前面需要跑完seamcut部分，根据seam两个输入画面各自计算mask，然后A B 分别是两个输入经过mask之后的画面
应该是padding不够导致顶点小于零了，另外，seamcut现在有点问题，效果不太正常
SeamCut应该是有点问题，时间久了我也没有精力去重新debug（应该是有bug）。学生时代写的代码可读性太差了，我自己也看不下去了，所以就不想再去维护了
可能没有了。实现起来倒也不难。跟踪的轨迹已经有了
Did you compile the seamcut project? any error message?
Well since you can compile it, why not try run it in Visual Studio or any other debugger... It may be a bug, or wrong filename...
Just stitch them one by one, there is no other option yet.