Planckian jitter: enhancing the color quality of self-supervised visual representations

Several recent works on self-supervised learning are trained by mapping different augmentations of the same image to the same feature representation. The set of used data augmentations is of crucial importance for the quality of the learned feature representation. We analyze how the traditionally used color jitter negatively impacts the quality of the color features in the learned feature representation. To address this problem, we replace this module with physics-based color augmentation, called Planckian jitter, which creates realistic variations in chromaticity, producing a model robust to llumination changes that can be commonly observed in real life, while maintaining the ability to discriminate the image content based on color information. We further improve the performance by introducing a latent space combination of color-sensitive and non-color-sensitive features. These are found to be complementary and the combination leads to large absolute performance gains over the default data augmentation on color classification tasks, including on Flowers-102 (+15%), Cub200 (+11%), VegFru (+15%), and T1K+ (+12%). Finally, we present a color sensitivity analysis to document the impact of different training methods on the model neurons and we show that the performance of the learned features is robust with respect to illuminant variations.