CVPR Daily - Tuesday

The team use optical flow techniques and PWC-Net architecture to keep the balance between efficiency and accuracy. Ping tells us that the work currently has a limitation. The method is based on forward warping . Given two frames, it forward warps the pixels via optical flow into the intermediate locations. This action causes problems because multiple pixels can map to the same location resulting in overlap, and the output is subject to holes where no pixels are projected to certain places in the intermediate frames. “ Even though our method results in very few holes, there are still some, and the next step is to design a mechanism to deal with this further, ” Ping advises. “ We rely on the quality of the underlying optical flow . If the quality of the optical flow is low in certain areas, the generated frames may also have low quality in those areas. We need to think about how to address these issues efficiently. ” Ping completed this work while he was interning at Adobe Research . His mentor there was Simon Niklaus , whom he says helped him immensely in this process: “ He’s an expert in this area working on some realistic applications of these techniques, so the advice he gave me in practical applications was invaluable. ” To learn more about Ping’s work [ID 2349], come to poster session 1.2 today at 14:30. 10 DAILY CVPR Tuesday Poster Presentation Visualization of forward warping via many-to-one (M2O) splatting and many-to-many (M2M) splatting. (a) With one source frame, M2M splatting suffers less from banding artifacts and provides improved robustness to ambiguities near the boundaries of discontinuous motion. (b) Banding artifacts can be alleviated with multiple source frames, yet M2O splatting still suffers from stray effects at boundaries due to its image formation model that is less flexible than M2M splatting.