InfraParis

We present InfraParis, an extensive dataset designed for autonomous driving applications. This dataset is characterized by its multimodal and multitasking nature, encompassing a total of 7,301 data samples. Specifically, it consists of 6,567 images in the training set, 189 in the validation set, and 571 in the test set. InfraParis is structured to support various tasks across three distinct modalities: RGB, depth, and infrared. These tasks encompass object detection, semantic segmentation, and depth prediction.

InfraParis encompasses four distinct tasks, each catering to different aspects of autonomous driving:

• Semantic Segmentation Task: This task involves images with associated semantic segmentation labels.
• Object Detection Task: In this task, the dataset provides images with bounding boxes delineating pedestrians.
• Supervised Depth Prediction Task: For this task, the dataset includes images with ground truth depth information.
• Unsupervised Depth Prediction Task: Here, the dataset comprises sequences of images along with ground truth depth data.

InfraParis has 4 modalities:

• RGB Images
• Infrared Images
• RGB Video
• Depth Images

Overview of annotated classes:

There are a total of 155 fine-grained classes, which are also aggregated to facilitate the use along with other datasets, e.g. Citiscapes: