Figures & data
Figure 1. Demonstration of images in specular highlight removal. Top: highlight images, bottom: diffuse images. We need to restore the diffuse image from the highlight image.
![Figure 1. Demonstration of images in specular highlight removal. Top: highlight images, bottom: diffuse images. We need to restore the diffuse image from the highlight image.](/cms/asset/77e50460-4175-44db-9a03-f9d928da293d/uaai_a_1988441_f0001_b.gif)
Figure 2. The image-to-image translation in specular highlight removal. The left is the highlight domain, and the right is the diffuse domain.
![Figure 2. The image-to-image translation in specular highlight removal. The left is the highlight domain, and the right is the diffuse domain.](/cms/asset/81178086-d735-4813-ba7e-42574bc426c5/uaai_a_1988441_f0002_b.gif)
Figure 3. The overall structure of our network. The generator consists of an attention module and autoencoder with skip connections. The discriminator is formed by a series of convolution layers.
![Figure 3. The overall structure of our network. The generator consists of an attention module and autoencoder with skip connections. The discriminator is formed by a series of convolution layers.](/cms/asset/9a97e205-f2ce-41fd-8876-80f2be35c862/uaai_a_1988441_f0003_oc.jpg)
Figure 4. Sample of highlight intensity mask. From left to right are highlight image, highlight intensity mask, and diffuse image which is ground truth.
![Figure 4. Sample of highlight intensity mask. From left to right are highlight image, highlight intensity mask, and diffuse image which is ground truth.](/cms/asset/1c6ac9b0-1a8e-43be-9669-123167fe8fd1/uaai_a_1988441_f0004_b.gif)
Figure 7. The structure of the pixel discriminator. Conv2D (a, b) means the input channel of the convolution layer is a, and the output channel is b.
![Figure 7. The structure of the pixel discriminator. Conv2D (a, b) means the input channel of the convolution layer is a, and the output channel is b.](/cms/asset/9c49f191-ec54-42d9-8278-478b24e521df/uaai_a_1988441_f0007_oc.jpg)
Figure 8. Samples from highlight dataset with 5 different view angles. The top row is highlight images, and the bottom row is diffuse images.
![Figure 8. Samples from highlight dataset with 5 different view angles. The top row is highlight images, and the bottom row is diffuse images.](/cms/asset/b527a87a-1bdd-43a5-b545-1c51bb908b35/uaai_a_1988441_f0008_b.gif)
Figure 10. Visualization of the highlight intensity mask generated by our attention module at different training steps.
![Figure 10. Visualization of the highlight intensity mask generated by our attention module at different training steps.](/cms/asset/39393835-8da8-46ec-b4b6-efcb8310fbf1/uaai_a_1988441_f0010_oc.jpg)
Table 1. The quantitative evaluation result of all test images
Table 2. The quantitative evaluation result of cases study