AdaInt: Learning Adaptive Intervals for 3D Lookup Tables in Real‑time Image Enhancement

The paper "AdaInt: Learning Adaptive Intervals for 3D Lookup Tables on Real‑time Image Enhancement" was accepted by CVPR 2022 and the code and models have been released publicly.

CVPR, one of the three top conferences in computer vision, received 8,161 submissions this year and accepted 2,067 papers (≈25% acceptance rate).

Color enhancement is a fundamental image‑processing task, but existing deep‑learning methods that embed a uniform 3D lookup table (3D LUT) suffer from high computational cost and limited non‑linear modeling ability, especially for ultra‑high‑resolution images (e.g., 4K).

To address these issues, the authors propose AdaInt, an adaptive‑interval learning framework. A lightweight convolutional network predicts non‑uniform sampling coordinates in the 3D color space and a set of basis 3D LUTs. The predicted coordinates are combined via a novel differentiable operator called AiLUT‑Transform, which uses a low‑complexity binary search to locate the appropriate grid cell and provides gradients for end‑to‑end training.

Experiments on the public FiveK and PPR10K datasets show that AdaInt improves objective metrics (e.g., PSNR) while adding negligible parameters and maintaining real‑time inference speed, outperforming existing state‑of‑the‑art methods.

In summary, AdaInt introduces image‑adaptive sampling intervals for 3D LUTs, achieving superior performance and efficiency in real‑time color enhancement, and the adaptive‑sampling idea may inspire improvements in other transformation‑heavy tasks.