AI Paper Atlas

Research map / Image generation / Image restoration

Colorful Image Colorization

Richard Zhang, Phillip Isola, Alexei A. Efros

ECCV 2016 · Image generation › Image restoration · approach: Image harmonization and color transformation · 3,424 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add class rebalancing to the classification loss (ablation)Ours (class) · ImageNet test (10k images)AuC rebalanced (%): 65.1 -> 67.3Improved
add class rebalancing to the classification loss (ablation)Ours (class) · ImageNet test (10k images)AMT labeled real (%): 25.2 -> 32.3Improved
add class rebalancing to the classification loss (ablation)Ours (class) · ImageNet test (10k images)AuC non-rebalanced (%): 91.6 -> 89.5Hurt
use classification loss instead of L2 regression, both trained from scratch (ablation)Ours (L2) · ImageNet test (10k images)AuC rebalanced (%): 64.4 -> 65.1Improved
use classification loss with rebalancing instead of L2 regression (ablation)Ours (L2) · ImageNet test (10k images)AMT labeled real (%): 21.2 -> 32.3Improved
fine-tune L2 network from the classification network instead of from scratch (ablation)Ours (L2) · ImageNet test (10k images)AuC rebalanced (%): 64.4 -> 66.2Improved
recolorize grayscale images with the method before VGG classificationVGG classifier · ImageNet test (10k images)VGG top-1 class accuracy (%): 52.7 -> 56.0Improved
fine-tune VGG on grayscale inputs instead of recolorizingVGG classifier · ImageNet test (10k images)VGG top-1 class accuracy (%): 56.0 -> 63.5Improved
add color to the input of the VGG classifier (ground truth vs grayscale)VGG classifier · ImageNet test (10k images)VGG top-1 class accuracy (%): 52.7 -> 68.3Improved
our full method versus Larsson et al. on AMT realismOurs (full) · ImageNet test (10k images)AMT labeled real (%): 27.2 -> 32.3No clear effect
train on SUN-subset comparison: our ImageNet model versus Deshpande et al.Ours (full) · SUN subset (Deshpande 240 images)AMT labeled real (%): 9.8 -> 17.2Improved
use color input (Lab) in conv1 instead of grayscale for PASCAL segmentationOurs · PASCAL VOC 2012 segmentationmIoU (%): 35.0 -> 35.6Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy