To ensure the most accurate object detection, you can set several parameters in the configuration file.
You can set the ClutteredBackground
parameter to specify whether the object, if present, occupies a large or small fraction of the image. The ClutteredBackground
parameter defaults to True
, specifying that an object might occupy only a small proportion of the image, both in terms of area and in terms of interest points. This setting allows Image Server to detect, for example, a small logo in a large, cluttered image. For many applications however, the object, if present, occupies a large fraction (>~25%) of the area. For example, matching a photo of a painting against a database of stolen paintings, or matching a logo against a database of logos to detect trademark infringement. The default settings still work for these scenarios, but they are sub-optimal. Specifically, the false positive rate is higher, and the algorithm runs more slowly and has greater sensitivity to lighting conditions. For this type of application, set ClutteredBackground
to False
.
You can set the Perspective
parameter to specify whether the object might appear at an angle, introducing perspective effects or ‘skew’ to the image. The Perspective
parameter defaults to True
, specifying that the object might be viewed at an angle. This default setting is the most generally applicable, and is normally the correct setting when the object is being viewed through a camera (for example, the board of sponsors’ logos that often appears in post-match interviews at sports events). However, for certain applications a 2-D geometry is more appropriate; for example, to detect a printed logo in a text document or scanned picture, or a broadcaster’s logo overlaid onto a video stream. For these cases, you can achieve much better false positive rates by setting Perspective
to False
. Image Server still allows for scaling, translation, and rotation in the plane, but does not allow tilting of the camera axis.
Going into more detail, Perspective=True
implies a homography between the image and the model, which allows for arbitrary focal length and pixel aspect ratio but does require the camera to behave approximately as a pinhole camera (the small distortions present in most cameras should not usually cause a problem). Setting Perspective
to False
implies a similarity transform.
|