Public training data for vehicle detectors in computer vision?

Question

This question is to anyone familiar with object (specifically vehicle) detection research.

I'm new to computer vision and am confused about training object detection classifiers. Specifically, the objective is vehicle detection. I've been reading through vehicle detection literature for weeks now, but I'm still a bit confused.

What I'm confused about is evaluation. For the evaluation of a system, the research community usually has a benchmarked dataset which can be used for testing data. But the performance of a system also depends very much on the data that was used to train it, no?

So aren't there any training datasets out there, too? That would make for far more uniform method comparisons. I seem to keep finding papers using benchmarked datasets for evaluation, but making no mention of where they got their training data from.

Answer 1

You typically use the same data set (or, rather, parts of it) for training and testing. I. e. you split the data set into a training set and a test set. A common technique for evaluating classifiers in general is called 10-fold cross validation. You split your data set 10 different ways such that 90% of the data is used for training and 10% of the data is used for testing. This way you get 10 different accuracy results, which you can then use to do statistical significance testing to show that your classifier is better than somebody else's.