From left to right, they are input, specular residue, diffuse, and tone-corrected diffuse images, respectively. The main reason for is that it allows to be trained with four-tuples image grops of the SHIQ and PSD datasets. Please download our SSHR dataset and see it for more details.

Our lighting collection is designed with Northern living in mind, by some of the worlds most talented designers, and produced using the highest quality materials.

For seven-tuples image groups (i.e. including additional albedo and shading), their index structure has the following forms:

Towards High-Quality Specular Highlight Removal by Leveraging Large-scale Synthetic Data Gang Fu, Qing Zhang, Lei Zhu, Chunxia Xiao, and Ping Li In ICCV 23

The following figure presents the pipeline of our three-stage framework. It consists of three stages: (i) physics-based specular highlight removal; (ii) specular-free refinement; and (iii) tone correction. Specifically, in the first stage (see (a)), we decompose an input image into its albedo and shading using two encoder-decoder networks ($E_a-D_a$ for albedo, and $E_s-D_s$ for shading). Then, the specular-free image can be estimated by multiplying the albedo and shading. In the second stage (see (b)), we feed the coarse result along with the input into an encoder-decoder network ($E_r-D_r$) to further refine it to alleviate visual artifacts. In the third stage (see (c)), we feed the refined result along with the input and its specular residue image into an encoder-decoder network ($E_c-D_c$) to adjust its tone so that it has the similar tone as the input as much as possible.

From left to right, they are input, albedo, shading, specular residue, diffuse, tone-corrected diffuse, and object mask images, respectively.

In this work, we propose a three-stage specular highlight removal network. To support network training and quantitative evaluation, we also present a large-scale synthetic dataset.