1. Video portrait segmentation
Reference from: https://zhuanlan.zhihu.com/p/48080465
video portraits can be seen as a special division of Semantic Segmentation task, he needs to put the video in each pixel of each frame are marked with a label category. Typical tasks and Semantic Segmentation There are two big differences: 1 category only move two (portrait and background) 2. Video characters is continuous, usually between several consecutive frames are very similar to the Mask of information Portrait we divided image at the i-th frame, i-1 can refer to the first frame has been considered a good Mask information.
2. Common Network
Currently doing some common neural network segmentation There are several networks can be selected as follows:
- MaskRCNN
- FCN
- UNET
- bless
- Tiramisu
but with unet network should be difficult to achieve real-time
3. Huawei desire
HUAWEI_HiAI
upgrade message video capture technology IT Home 1 November spawned a large number of high-quality video, but current technology is still difficult to support the many creative shooting. Because of space limitations and props, a lot of special effects need to spend a lot of time and energy to green screen filming and post-production, resulting in high costs shoot. Recently, HUAWEI HiAI line has the ability to split video portraits, one-click access to help developers implement video background processing and rendering powerful figures on Huawei's Developer Connection.
