L-CoIns: Language-based Colorization With Instance Awareness

The right woman is dressed in shirt of violet color. L-CoDe L-CoDer Ours ML2018 The middle woman is dressed in orange and the right woman in yellow. L-CoDe L-CoDer Ours ML2018 Three women are dressed in pink. L-CoDe L-CoDer Ours ML2018 The middle woman is dressed in yellow, the left woman in blue, the right woman in red. L-CoDe L-CoDer Ours ML2018 Grayscale Figure 1. Language-based colorization results given four different language descriptions, compared with ML2018 [29], L-CoDe [42], and L-CoDer [6]. Top left: For the description that has clear correspondences between color words and object words, our method correctly colorizes all corresponding regions. Top right: For the description that assigns distinct colors for every instance corresponding to the same object words, our model predicts the exact correspondence between the instance region and the color word. Botton left: For the description that includes unobserved correspondences between color words and object words, our method could adaptively parse the sentence and determine the correct semantics for colorization. Bottom right: For the description that is against the statistical correlation between luminance and color words, our method shows the robustness and colorize description-consistent results.