0 Preface
Since 2012 Alexnet turned out to deep learning on the identification and segmentation of high-level semantic features great success. However, the depth of learning data based on a very large number of marked image level or stage of the bounding box, there are many limitations in practical applications.
1 are denoted by a rough, many target itself is not rectangular, a square marked introduce errors transactions;
2, out of the bounding box labeled training framework divided accuracy is not reached superhuman precision, because the label data itself is not reached superhuman level of accuracy;
3, to solve the occlusion problem is very difficult;
For these reasons, it was put forward new ideas pixel level marked an attempt to solve the above problems.
1 pixel-level tagging
To link it
https://mp.weixin.qq.com/s/3lK1as-cdMxFVEtQX__dAw
Stage 2 marked voxels