文章目录
- 一、基本信息
- 1.1 论文基本信息
- 1.2 课程基本信息
- 1.3 博文基本信息
- 二、论文评述(中英双语)
- 2.1 研究问题(Research Problem)
- 2.2 创新点(Innovation/Contribution)
- 2.3 优点(Why this paper is well written)
- 2.4 不足(Inadequacies)
一、基本信息
1.1 论文基本信息
标题:Neural Layered BRDFs
来源:ACM SIGGRAPH 2022(中国计算机学会CCF推荐国际学术会议-计算机图形学与多媒体-A类)
作者单位:南京理工大学、南开大学、Adobe Research、加利福尼亚大学
原文:https://dl.acm.org/doi/10.1145/3528233.3530732
小组报告时间:2023年上半年(本人为小组组员)
1.2 课程基本信息
课程名称:科技论文写作
开课单位:浙江工业大学计算机学院
课程性质:硕士课程-专业课-核心课程
先修课程:硕士英语、机器学习等
教学目的:使学生在学习了专业课程,并经历了一定的科研项目试验过程的基础上,了解科技论文的写作目的、掌握其写作的基本过程和规则,从而提高研究生的科技论文写作效率。
课程思政元素:实事求是、精益求精、突破陈规。本课程鼓励学生写论文要符合三真,即真问题、真方法、真数据,以实事求是的态度撰写科技论文;同时写作的过程中,对所提的每个观点进行自我发问,做到精益求精;最后还要对所提问题和方法进行深入思考,是否能够突破陈规,体现创新性。
课程大纲:
- 研究问题确认
1.1 讲逻辑要先区分事实与观点
1.2 研究是什么,研究者如何看待
1.3 从研究主题到具体问题
1.4 找到有用的文献
1.5 与文献交互- 如何草拟研究论文初稿
2.1 规划论文思路
2.2 论文初稿设计
2.3 草拟论文
2.4 论文写作工具- 如何修订论文
3.1 如何选择表格和图形呈现论据
3.2 草稿的修订
3.3 拟出最终的引言和结论
3.4 修订句子
3.5 从论文评语中学习- 如何做报告与科研精神
4.1 设计口头报告
4.2 口头报告要适合聆听
4.3 关于科研精神
教学参考资料:《芝加哥大学论文写作指南》
考核方式:课内考察,采取课堂讲授、专题学术报告、讨论、课程报告相结合的教学方式
1.3 博文基本信息
本课程要求学生组成小组,在选择的研究领域中,选择来自顶会顶刊的论文进行阅读。小组内的每位成员必须对论文发表自己的见解,向组长提交一份书面报告,由组长总结所有小组成员的观点,使用PPT+口头报告的方式进行展示。
本小组选择的研究领域为计算机图形学(Computer Graphics,CG),它是一种使用数学算法将二维或三维图形转化为计算机表示的科学。其主要研究内容是:如何在计算机中表示图形,进而利用计算机进行图形的计算、处理和显示。计算机图形学的核心目标在于创建有效的视觉交流,在科学、娱乐等领域和艺术创作、商业广告、产品设计等行业中发挥着重要作用。
以下部分来自博主个人的书面报告,该报告形成过程中经过小组交流讨论,但由于本人研究方向并不是计算机图形学,实际上是以大同行的视角对论文进行评述,一家之言,仅供参考。以下部分为终稿,若无特殊情况,将不再进行修改。
二、论文评述(中英双语)
2.1 研究问题(Research Problem)
在计算机图形学中,双向反射分布函数(BRDF)被广泛用于表示和渲染多层材料。然而,现有评估方法存在高方差、高成本、精确性低等问题。
In computer graphics, Bidirectional Reflectance Distribution Functions (BRDFs) are pervasively used to represent and render layered materials. However, existing methods have the limitations of high variance, high cost and less accuracy.
2.2 创新点(Innovation/Contribution)
作者提出用神经网络将BRDF压缩为潜在表示,在神经空间中进行分层,并通过分层网络对这些潜在向量执行学习的分层操作。与最先进的方法相比,本文提出的BRDF评估方法具有无噪声和计算效率高的特点。
The authors proposed to perform layering in the neural space by compressing BRDFs into latent codes via a proposed representation neural network and performing a learned layering operation on these latent vectors via a layering network. The proposed method is noise-free and computationally efficient compared to the state-of-the-art approach.
2.3 优点(Why this paper is well written)
(1)摘要部分简练而全面,覆盖全文要点:介绍、方法、实验结果、结论。(The abstract is terse while comprehensive, covering the full text points: introduction, methods, experiments and conclusion. )
(2)文章结构框架合理,第3节、第4节的小节标题与大节标题相互对应。(The paper has a reasonable structure or framework. For instance, the subtitles of the third and forth section correspond to the titles of these sections.)
(3)引言部分逻辑清晰。第一段简明扼要地介绍了研究问题的背景,以从一般到特定的顺序明确主题,并举例说明应用场景;第二段和第三段分别对解决当前问题的旧方法和解决更简单问题的新方法进行了陈述和评价;在第四段提出自己的主张,声明了本文的贡献,为读者提供了清晰的导向。(The logic of the introduction is clear. The first paragraph introduces the background briefly, identifies the topic in a way from general to specific, and puts up several typical application scenarios. The second and third paragraph state and evaluate the old methods of current problems and the new methods of simpler problems, respectively. The fourth paragraph offers contribution claims, providing a clear guidance for the reader.)
(4)在相关工作部分,按照多个类别进行组织。对前人的工作进行了充分论述,介绍了研究的来龙去脉,比较他们的差异并进行归类,并与本文的方法比较,突出本文的贡献。其中特别提到了当前的真相方法,通过声明本文的方法接近真相,为本文方法的有效性提供了逻辑上的有力支持。(The related work is organized in several categories. Previous work is fully discussed, the context of the study is presented, their differences are compared and used for classification. The contribution of the article is highlighted by comparing with these classified methods. Among them, the ground-truth is particularly mentioned, providing a logically strong support for the effectiveness of the proposed method. )
(5)在第三节的开头,用一段话简要介绍了这一节的内容。在3.1节用公式对问题进行了描述。这里的formulate用得非常准确,对问题精确的形式化定义是解决问题的第一步。在这一部分,首先确定了问题的范围,接着交代了本文的核心概念BRDF与相近概念的关系。在3.2节,用图2表示了评估网络的详细架构。使用相连的三角形和梯形,巧妙地表示了网络的各个组件,节省了空间。在表 1 中将本文的方法与三项相关工作进行了比较,突出了本文方法的特点。(In the beginning of the third section, a paragraph is used to introduce the content of this section briefly. The problem is described in 3.1 by equations. The word “formulation” is used rather accurately, as the precise formal definition of the problem is the first step in solving it. In this part, the scope of the problem is first determined, then the relationship between the core concept of the article (i.e., BRDF) and similar concepts is explained. In 3.2, Figure 2 shows the detailed structure of the evaluation network. Closely connected triangles and trapezoids are used to represent the components of the network skillfully, saving the space for typesetting. Table 1 compares the proposed method with three related works, emphasizing the feature of the proposed method.)
(6)在实验部分,使用多种材料组合成多种分层材质进行神经网络的训练,数据量充足。(In experiments, multiple kinds of materials are used to generate 12720 layered BRDFs to train the networks, which guarantees for sufficient data.)
(7)在结论部分,回顾了主要贡献,明确了当前方法的限制,并对未来工作进行了展望。(In conclusion, the main contribution is recapped, the limitations of proposed method are cleared, and the future work is discussed.)
(8)提供了补充资料,显著提高了论文的可读性。(The supplementary material is provided, which significantly improves the readability of the paper.)
(9)引用的参考文献较新:在19篇引用文献中,有14篇在近5年发表。(Relatively new references: 14 of the 19 items were published in the last 5 years.)
(10)全文的过渡词使用恰到好处,衔接自然,过渡流畅。配图美观清晰,赏心悦目。(The transition words in the paper are properly used, results in natural connection and smooth transition. The figures are artistic and clear, which bring pleasant experience to the readers.)
2.4 不足(Inadequacies)
(1)摘要最后一句话中的“神经代数”(neural algebra)在正文中只是一笔带过,有博眼球的嫌疑。(The word “neural algebra” in the last sentence of the abstract is simply described in the main body, which is a suspicion of attracting eyeballs. )
(2)对于公式2,没有说明Nlayering和Vlayered的含义。(As for Eq.2, the meaning of symbol Nlayering and Vlayered are not explained.)
(3)3.2节中,对于为什么要离散化输入BRDF,可补充说明。(In 3.2, the necessity for the discretization of the input BRDF could be complemented.)
(4)没有公开代码,而且没有用伪代码进行描述。(The code is not released, and pseudocodes are not used.)