Select Images for Training
The following table describes the training requirements for 2-D and 3-D objects:
Object type | Requirements |
---|---|
2-D |
One or more training images. Media Server creates a separate model for each training image. For example, if a company has many different variations of its logo, you only need to add one 2-D object to your object database. Add all of the training images for the logo to the same object, even if some variations of the logo do not look like the others. |
3-D |
Multiple training images depicting the same 3-D object from all angles. Provide images of the object from all angles that you expect it to appear in target images. Micro Focus recommends that the images are taken in one continuous process. One way to obtain training images is by recording a video of the object from all angles and extracting images from the video. As a rough estimate, between 20 and 40 images of an object are usually sufficient to provide accurate detection in a small image. Media Server uses all of the training images to create a single model. |
A good training image for an object:
- contains only the object, with as little background as possible. The object should be the dominant feature in the image but there should be at least 16 pixels of background surrounding the object. If any parts of an object touch the image edges, the features corresponding to those parts of the image are likely to be lost.
-
includes transparency information, if you intend to recognize the object against many different backgrounds (for example, superimposed on TV footage). Ideally all parts of the image that are not part of the object are transparent (by having their alpha channel set appropriately).
NOTE: Only some image formats support transparency information, for example .PNG and .TIFF. Media Server does not support transparency information in .GIF files.
-
has not been converted from another file format to JPEG. Converting images to JPEG introduces compression artifacts. If you have JPEG images, avoid making many rounds of edits. Each time that a new version of the image is saved as a JPEG, the image quality degrades.