1. Ours Our method trained with class rebalancing (λ =
1
2
in Equation 6).
2. Ours fine-tuned without class-rebalancing To specifically test the contribution
of class-rebalancing, we finetune our trained model on a nonrebalanced
objective (λ = 1 in Equation 6) for 16k iterations.
3. Dahl [2] Regression model using a Laplacian pyramid on VGG features.
4. Random As a naive baseline, we copy the colors from a randomly chosen
image from the training set. This baseline has natural image color statistics.
5. Gray We use a naive baseline which colors every pixel gray