解读Towards Unified Depth and Semantic Prediction from a Single Image（3）
上一次讨论的最后，补充了一些条件随机场 (CRF) 的知识，还不知道够不够用
知识嘛，活的东西，不够再补，活学活用
作者总结了自己方法的三大优势：
1, We propose a unified framework for joint depth and semantic prediction from a single image
2, We formulate the problem in a two-layer HCRF to enforce synergy between global and local
predictions, where the global layouts are used to guide the local predictions and reduce local
ambiguities, while the local results provide detailed region structures and boundaries
3, Through extensive evaluation, we demonstrate that jointly addressing the two problems in
our framework benefits both tasks, and achieves the state-of-the-art
接下来进入Related work这一部分，
作者开篇是这样讲的，
The literature of depth estimation and semantic segmentation is very rich when considering
them as two independent tasks
但是呢，
Interestingly, though developed separately, the techniques used to solve the two tasks are quite
similar
很多学者致力于这样的研究，于是呢，
Inspired by these work, we also use CNN to train our model for joint global and local prediction
但是有一点，
While promising, the joint inference of these two tasks to enforce consistency between them is
an under-explored direction in the literature
具体的细节我没有讲述，可以直接看原文，
不管怎样，一个思想并不是凭空诞生的，没有前人辛苦的灌水，后人难有新的发现
读完了Section 1，开始阅读Section 2 Formulation
作者先说明使用HCRF的背景和原因，
1, Semantic inference should consider both short-range pixel-wise interactions and high-order
context
2, The consistency in depth and semantic prediction should also be enforced both globally and locally
3, To this end, instead of a standard pixel-wise Conditional Random Field, we propose a two-layer
Hierarchical Conditional Random Field (HCRF) to formulate the joint depth and semantic prediction
problem
下次接着分析后面的！
