Friday, October 24, 2014

Nvidia FlexISP Promises Image Quality Improvements

Nvidia Research publishes its paper and presentation "FlexISP: A Flexible Camera Image Processing Framework" by Felix Heide (UBC), Markus Steinberger (TU Graz), Yun-Ta Tsai (NVIDIA), Mushfiqur Rouf (UBC), Dawid Pająk (NVIDIA), Dikpal Reddy (NVIDIA), Orazio Gallo (NVIDIA), Jing Liu (UC Santa Cruz), Wolfgang Heidrich (KAUST), Karen Egiazarian (TUT), Jan Kautz (NVIDIA), Kari Pulli (NVIDIA), to be presented at SIGGRAPH Asia in December 2014.

While conventional ISP pipelines are usually defined as a series of cascaded modules, each responsible for addressing a particular problem, FlexISP promises to achieve everything in a single step. Vimeo video explains:


  1. Well - when you have a poor noise model and crappy prior generation ISP, flexISP looks better. NVidia has never managed to impress anyone with their imaging. As one of the first major phone/tablet OEMs to adopt NV chipsets, we decided never again after the first generation device. NV swore by their 8-bit ISP and would give us funny looks when we insisted on 10/12-bit color pipes. We never managed to avoid bleeding and color artifacts without seriously degrading the overall image quality using their ISP. How could a company that understands GPUs so well so totally screw-up imaging?

  2. It is actually a very interesting paper, however I think the abstract in most reports will be misleading (including this one).

    The paper tries to show that many goals of different image applications can be all covered by a general optimization framework. Therefore you don't have to design the pipeline for each specific applications. However, it is definitely not a "single-step" processing. The particular optimization solver used in the paper requires multiple iterations to converge, and inside each iteration some steps also requires additional iterations, or perform operations very similar to some blocks in the existing pipelines. Overall its complexity can be higher than many ISP.

    I take the advantage of this method is quick prototyping for new applications. As long as you can find a way to formulate your image capture process into the proposed framework (which seems to be very easy to me), the image quality of the output can be comparable to some state-of-the-art (in the public domain). So you can focus on exploring application-level design and leave the ISP design later. This can be a very useful tool in that sense.

  3. ok SIGGRAPH,and AMD(ati) please common (^_^)


All comments are moderated to avoid spam.