In this paper we provide a method for evaluating interest
point detectors independently of image descriptors.
This is possible because we have compiled a unique data
set enabling us to determine if common interest points are
found. The data contains 60 scenes of a wide range of object
types, and for each scene we have 119 precisely located
camera positions obtained from a camera mounted on an industrial
robot arm. The scene surfaces have been scanned
using structured light, providing precise 3D ground truth.
We have investigated a number of the most popular interest
point detectors. This is done in relation to the number
of interest points, the recall rate as a function of camera
position and light variation, and the sensitivity relative to
model parameter change. The overall conclusion is that the
Harris corner detector has a very high recall rate, but is
sensitive to change in scale. The Hessian corners perform
overall well followed by MSER (Maximally Stable Extremal
Regions), whereas the FAST corner detector, IBR (Intensity
Based Regions) and EBR (Edge Based Regions) performs
poorly. Furthermore, the repeatability of the corner detectors
is quite unaffected by the parameter setting, and only
the number of interest points change.