Vehicle Environment Dataset

We collected the data from ten drivers that drove vehicles in St. Petersburg, Russia, for a few months. The image size from the road cameras is 480x640. For training, we scaled them down to 420x420 and applied center cropping to a 416x416 area to exclude border areas that contain  noise and distortion. To annotate the data, we used pseudo-labeling and an ensemble of different open-source models that were trained on the KITTI dataset. The ensemble is based on a weighted mean average using the following weights [0.4, 0.3, 0.2, 0.1] from top to bottom. For each image in the dataset, we used the following four models (LapDept, DTP Hybrid, BTS, and VNL) to obtain the predictions. After that, we took the weighted average between all four masks and saved the results as our ground truth.