Texture compression is widely used in real-time rendering to reduce storage and bandwidth requirements. Recent research in compression algorithms has explored both reduced fixed bit rate and variable bit rate algorithms. The results are evaluated at the individual texture level using Mean Square Error, Peak Signal-to-Noise Ratio, or visual image inspection. We argue this is the wrong evaluation approach. Compression artifacts in individual textures are likely visually masked in final rendered images and this masking is not accounted for when evaluating individual textures. This masking comes from both mapping textures onto models and using different textures on the same model such as diffuse, gloss and bump maps. We propose to evaluate final rendered images using rigorous perceptual error metrics. Our method samples the space of viewpoints in a scene,renders the scene from each viewpoint using variations of compressed textures, and then compares each render to a ground truth render using uncompressed textures from the same viewpoint. We show that masking has a significant effect on final rendered image quality and that GPU compression algorithms are too conservative and reduced bit rates are possible while maintaining quality.
Proceedings Title: Proceedings of the ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games
Conference Dates: March 14-16, 2014
Conference Location: San Francisco, CA
Pub Type: Conferences
MIP mapping, bump maps, texture compression, image quality assessment