OpenCV comes with several predefined structures, including CvScalar, CvPoint, etc. But the three most important structures are CvArr, CvMat and IplImage. You can call them the primitive data types for OpenCV. From the C point of view, these are not exactly 'primitive' data types. But they are the fundamental data types on which OpenCV manipulates (resize, threshold, etc).

Inheritance

The three structures are "related" to each other by "inheritance". No, its not the C++ or Java inheritance. OpenCV was made using C, so there's no such concept. But, their relation mimics inheritance.

CvArr is the "base" class. Think of it like an abstract class. It is merely used as parameters for various functions. You never actually create an "object" of CvArr.

CvMat is "derived" from CvArr. Logically, a matrix is nothing but an extension of an array. So it makes sense. These matrices aren't the ones you had studied in school. Its a generalization. Matrices can be N dimensional. And each "element" of this matrix can hold more than one values (technically, every element is a tuple).

IplImage is the structure that actually holds images that have to be manipulated. You'll use this a LOT in OpenCV. With this structure, you can store images in multiple formats: single, three, or four channels, and each channel holding integers (in various formats) or floating point decimals.

This is your "normal" image is a 3 channel 8-bit image. While manipulating, you'll come across several operators that return "images" in other formats as well.

Another important point to note: If you see CvArr as a parameter to some function, you can pass a CvMat or an IplImage. Both are acceptable. Similarly, if you see CvMat as a parameter, you can pass a CvMat (obviously) or an IplImage. But If you see IplImage, you can only pass IplImage.

It works just like inheritance, where you can pass derived classes in place to parent classes.