FLUID: Formulated LiqUid Instability Dataset

Maurizio De Micco, Fabio Zonfrilli, Massimiliano Maria Villone, Giovanni Poggi, Luisa Verdoliva, Vincenzo Guida

Abstract: We present FLUID, a new dataset of more than 47,000 pictures of softener samples developing typical instabilities of formulated liquids of industrial interest. The full-color pictures, labelled by a pool of experts, are organized in 5 categories (stable, phase separation, creaming, cracking, and flocculation) and can be accessed either as single frames or as an ordered series of 8 to 10 frames, equally spaced in time, belonging to the same sample. FLUID is publicly available on-line at the link https://github.com/mauriziodemiccounina/FLUID.git and is intended to become a benchmark dataset. It can be used to perform instability detection analysis and potentially forecast of failures in formulated liquids. Furthermore, we show that it represents a precious tool to train learning-based methods and boost their performance by means of deep architectures.