Solving a challenging task 【Neural Recon in the wild, SIGGRAPH】
Novelty
Novel task 【Polymorphics-GAN, CVPR 2022】
Novel pipeine 【Deep marching cubes, CVPR 2018】
Novel module 【ConvONet, ECCV 2020, spotlight】
Novel loss 【mipNeRF 360, CVPR 2022, oral】【LipMLP, SIGGRAPH2022】
Full paper
goog paper
The heatmap generated by class activation mapping [24] highlights regions specific to good papers, e.g., teaser figures in the first page for illustrating the main ideas, tables/plots showing a sense of thoroughness in experimental validation, impressive math equations, and arrays of colorful images for qualitative results from benchmark datasets.
bad paper
(Top) Failing to fll the paper into a full eight-page paper is a discriminativevisual cue for bad paner(Bottom) The generated heatmaps focus on the top-right corner of the first page. This suggests that the absenceof illustrative figures in the first two pages may cause the paper more difficult to understand.
Teaser
Result
Key idea
Comparision
Introduction
Treat readers as guests visiting your house
Related work
Method
overview figure
Include full architecture, loss functions, training + inference, (forward + backward)
Use math symbols + text, coherent with equations in the main paper
Detailed caption. Readers should be able to understand the method after reading the figure + caption
content
Pyramid Sructure
Start the section with an overview of the method
Start every paragraph with the key message(中心句放最最前面)
Same text cannot be repeated, but the same message can be repeated with different representations (such as equation, figure)
Implementation Details (mentioning important ones, more detailed information can be provided in supplementary)
Experiments
All claims should be supported
Always ask yourself whether a claim needs an experiment for support or not when writing introduction
content
4.0 - Datasets - Baseline -
4.1 comparision to SOAT
4.2 ablation study
Have a toy example to explain the method if possible
Ablation study, show quantity and quality results
Mention key messages in the figure/table caption
Write a python scripts for generating tables and figures. Do it early
Analyze the results! Why is your method better? why dose the baseline fail?\
每个表格表达一条信息
Supplementary
Most accepted papers come with supplementary materials
PDF, video (<3min), or both
Referring to supp. mat. in the main paper
Before submission
Read the paper multiple times before submission!
Proof-reading by colleagues
A good paper: whenever a question can be raised, the answer is provided in a few sentences
Use the same terminology for the same concept
Think from the perspective of the reviewers
Check paper for standard errors:
incorrect mathematics/reasoning/experiment: is this right?
omitted material (experiment/context/citations): s stuff missing?
done before: if so, by whom? why is this not different?