Skylight polarization phenomenon is at the origin of a recent growing interest for bio-inspired navigation. Skylightbased orientation sensors can be simulated on the basis of physical models. In parallel, machine learning algorithms require a large amount of data to be trained. However, while some simulated databases already exists in the literature, a public database composed of real-world color polarimetric images of the sky in various weather conditions does not. In this study, a long-term experimental device is presented, designed to be left in a distant roof to acquire data over several months, using a Division-of-Focal-Plane polarization imager with a fisheye lens mounted on a rotative telescope mount. An open-source mechanical and electrical design is proposed for easy replication at other locations, with an algorithm to get the sensor's orientation and geometrical calibration in the East-North-Up frame. A sample one-month long dataset is provided on a public archive.