Discover and read the best of Twitter Threads about #IngridDaubechies
Most recents (1)
SSIM has become a common loss function in computer vision. It is used to train monocular depth models for self-driving cars, invert GANs, and fit NeRF models to training images. The explosion of SSIM-based models raises a fundamental question: what the hell is SSIM? 🧵
SSIM measures the similarity between two images. Humans are insensitive to the absolute brightness/color of pixels, but very sensitive to the location of edges and textures. SSIM mimics human perception by focusing primarily on edge and textural similarities.
Here’s an example. The contrast adjustment between these two images of #IngridDaubechies makes them 20% different when measured using the 2-norm. But in the SSIM metric they are 98.5% similar (1.5% different). 
