Benchmarks

The Prague texture segmentation data-generator and benchmark is a web based (http://mosaic.utia.cas.cz) service designed to mutually compare and rank different static or dynamic texture segmenters, and to support new supervised or unsupervised classification methods development. The benchmark verifies their performance characteristics on monospectral, multispectral, bidirectional texture function (BTF), satellite or dynamic texture data and enables to test their noise robustness, scale, and rotation or illumination invariance.

Time-of-Flight (ToF) cameras use specialized sensors and modulated infrared light to simultaneously obtain depth, amplitude and intensity images. Depth images from such cameras suffer from various errors which exhibit a more complex behavior than traditional intensity images. Of these errors, the phenomenon of multi-reflection or multi-path interference poses the biggest challenge to researchers. It is caused by indirect light paths between camera and light source and is therefore dependent on scene geometry. While simulated data can be used for ground truth evaluation and whitebox testing, current simulators do not model multipath effects. The method we present is capa-ble of simulating all scene-dependant effects by taking global illumination into consideration. This is accomplished by modifying a bidirectional path tracing algorithm such that it takes the time-dependent propagation of modulated light in a scene into consideration. Furthermore, by combination of the proposed method with a previous hardware simulator we are capable of reproducing all effects in ToF cameras. The system was validated both on test targets with known real Time of Flight camera responses as well as qualitatively on a more complex room scene.

The task associated to this dataset is tracking multiple drosophila larvae. Such a tracking is required in the quest to elucidate the genetic basis of Drosophila's behaviour. This dataset was used in the article "Tracking indistinguishable translucent objects over time using weakly supervised structured learning". We provide the raw data, an intermediate segmentation of the foreground and the gold standard used in the evaluation of that tracking algorithm.