Computer Vision News - January 2022
6 Computer Vision Research High receptive field perceptual loss The focus of large mask inpainting is shifted towards understanding of global structure. Naive supervised losses require the generator to reconstruct the ground truth precisely. The visible parts of the image often do not contain enough information for the exact reconstruction of the masked part. In contrast, perceptual loss evaluates a distance between features extracted from the predicted and the target images by a pre-trained network. It does not require an exact reconstruction, allowing for variations in the reconstructed image. Adversarial loss Adversarial loss was used to ensure that inpainting model f θ (x′) generates naturally looking local details. A discriminator D ξ (•) works on a local patch-level. Only patches that intersect with the masked area get the “fake” label. The final loss function The final loss function is the weighted sum of the discussed losses. The last component of the system is a mask generation policy. The way the masks are generated greatly influences the final performance of the system and unlike the conventional practice, e.g. DeepFillv2, a strategy with aggressive large mask generation was used. Samples from polygonal chains were uniformly dilated by a high random width (wide masks) and rectangles of arbitrary aspect ratios (box masks). The outcome of the network was shown to outperform a range of strong baselines on low
Made with FlippingBook
RkJQdWJsaXNoZXIy NTc3NzU=