Learning Similarity Metrics for Dynamic Scene Segmentation Supplementary material

The data from the SynthDB dataset used for training consists of the 99 sequences featuring 2 textures [1]. Each texture is labeled as one of these 12 classes: grass, jellyfish, pond, boiling, escalator, fire, river-far, river, steam, plant-a, plant-i, and sea-far. All results reported on the SynthDB and Dyntex datasets used a single scale, i.e. S=1. In our experiments, the use of multiple scales did not have a significant impact on the results for this task (unlike with object and motion segmentation). We believe it is a consequence of the limited diversity of training data. Using multiple scales is more likely to be beneficial if the model was trained on scenes that include dynamic textures of varying spatial extent. The results on the Dyntex sequences reported in the main paper correspond to the last level of the segmentation, i.e. when only 2 segments are remaining. Practically, one segment correspond to the main dynamic texture of the scene, the other to the background (or more static elements, keeping in mind that the camera is moving in some sequences).