STEP stands for The Segmenting and Tracking Every Pixel benchmark, which includes 21 training sequences and 29 test sequences. This dataset is based on the KITTI Tracking Evaluation and Multi-Object Tracking and Segmentation (MOTS) benchmarks.

This dataset adds dense pixel segmentation labels for each pixel. In this benchmark, each pixel has a semantic label and all pixels belonging to the most salient object categories (cars and pedestrians) have a unique tracking ID.