node == process

启动了标明坐标系的launch文件后

后期可以避免重名

不需要依赖

后面三个是依赖

our probability matrix is defined here. https://ogunlao.github.io/blog/2020/07/17/breaking-down-ctc-loss.html#:~:text=This%20becomes%20our-,probability%20matrix,-%2C

e.g. "door" and given timesteps T: Up to t = 4. we might be still staying at d. Situation can be like: blank blank blank d

Note that simga is for charctor seperation. underscore is for word seperation

e.g. door --> "d" "blank" "o" == "d" "o" but the skip is permited because "d" and "o" are differnt on the either side of the a blank.

Now we are creating a grid to get the expanded target sequence.(a.k.a expanded output matrix)

e.g. the target seqence is "door" if the row doesn't include token "d" "o" "r", we remove it.

the sum overs paths corresponding to a specific labelling can be broken down into an iterative sum over paths corresponding to prefixes of that labelling.

We have many different paths to get the same result. loss is computed by taking the sum of the probaiblities of all these paths(th p of they occuring)

So far, we already know, given a [X,Y,Z], how do we map it into [u, v].

We use the kernel to compare the given pixel with its neighborhood.

How do you let the computer know there is an edge here?

作者更新得挺勤快