After you got some idea about vision, this time I’ll write about modern literature. Like programming languages, scientific papers look like English but actually written in a special technical language.
There is even a dictionary. Like programming languages, technical language doesn’t ensure correctness. Given the poor quality assurance these papers are going through, (only few readers, nobody checks the computer programs seriously), and given their complexity, it is surprising the number of mistakes is so small.
In computer vision, there are more than 1000 publications a year. That means one can spend all his time reading. It becomes harder to think about something new, or to assert that something is really new. And for every paper that gets published there is another which had been properly reviewed.
For those who are unfamiliar with the reviewing process, take a glimpse behind the scene. People in this area are very critical about papers, especially those written by others.
When a reviewer has a bad day every paper can be rejected. Objections can always be made regardless the content. If the paper is theoretical, what is the application? If it is technological, where is the new theory? If it is using an old idea, what is new? If it is a new idea, it wasn't tested enough, and there must be a good reason why people didn’t do it until now. If it uses an approximation or an assumption, it is easy to come up with counterexamples where they break. If it is deterministic, then how things behave under uncertainty? If it statistical it is questionable whether we can trust the conclusions. Generally, if the authors were able to summarize a solution in 10 pages, the problem wasn’t that hard to begin with. And what about the problems they didn’t solve in the paper? Finally, one can always claim a paper is uninteresting.
But there are ways to publish. Remember that the good authors and the bad reviewers are essentially the same people. The professionals know the marketing game. Reviewers don’t like to read too much. Although it is inappropriate to submit comics, good looking images and 3D figures make a difference. Some people attach videos and complete presentations as supplementary material. Reviewers are also sensitive to eigenvalues. If you can’t use eigenvalues, an integral, or better a double integral, is an ace. Otherwise just stuck a log somewhere. Another approach is to make it clear that a famous figure stands behind the paper. If conference organizers want him to attend, they must accept the paper.
You might be wondering what about the idea of the paper, the original contribution. A standard way around is to use the combinatorial method. To illustrate, suppose there are 10 relevant approaches. If we take their combinations, that makes 1024 possible papers (including the empty paper, which is as good as many other combinations). In the past people stuck to their area, but today a paper, say on wavelets in databases, doesn’t sound unnatural.
There are ways to automate the process, and even generate the whole paper. However, with 1000 publications a year it gets tougher.
It takes some experience to understand what authors really mean in these papers. There are papers written in very decisive tone and presented impressively. But after talking to the authors privately, you realize that the people who know best the problems with scientific papers are the same people who wrote them.
