A calibrated 3D dataset for automatic evaluation of keypoint detectors and descriptors

Previous evaluations of keypoint detectors and descriptors have tended to focus on planar scenes, relying
on homographies to generate the ground truth. Evaluations on 3D scenes include approaches using trifocal
tensors on natural scenes those using calibrated images from a turntable. In order to provide an
accurate evaluation of keypoint detectors methods on real 3D scenes, we needed a dataset that (a) contained sufficiently large objects so that 3D effects were noticeable, (b) the background did not distract from the object of interest and (c) the ground truth was available.  To overcome these problems, we introduced a new dataset, `Cambridge toy cars dataset’. This report describes the dataset and acquisition process.

Sadly, the dataset itself is now long lost, but the techniques described in here may be useful to someone.