[210112] Ross Girshick: writing good research papers
website: https://zhuanlan.zhihu.com/p/104279421
What did I learn from your paper?
- A paper should be about
a single focused idea or question
. - “Idea” usually means method. What should I learn?
- Under what conditions does it work?
- When does it work?
- If the idea has multiple components, which is the most important one?
- Which implementations are the most important?
- I seldom care if “your idea” + “unrelated ideas/tricks” --> S.O.T.A results
- my first priority: to learn something interesting about your ideas.
one table, one message
Ross 用他的 Mask R-CNN 为例,论文中有很多表格,每张表都只表达1个信息。
Support all of your claims
所有的 claim 应该用 citation 或者 experiment 来支持。否则,不要出现不合格的表达,比如:
- “Intuitively, increasing X is important for Y…”
This statement is your intuition (not fact), the reviewer may disagree! - “Increasing X may lead to improved Y…”
Expresses uncertainty or that some conditions may apply
Ross举了一个例子,在Mask R-CNN论文中,论文中说alignment非常重要,因此提出了 RoIAlign,为了支持这个观点,有一个实验结果表格来支撑。
speed/accuracy claims
与当前算法的比较经常是不可控的(uncontrolled)。主要表现为如下五个方面:
- 使用不同参数时,算法准确率不同
- 不同参数及硬件环境中,算法速度不同
- inference details 不同时,算法速度不同
- 有些人是以优化算法速度为生的!他们能做到10-100倍速度提升
- ?
因此,我们对于算法准确率及速度的提升保持怀疑态度(taken with a large grain of salt)。而我们在做实验时,注意下面四项即可: - Making training settings as similar as possible
- Making inference settings as similar as possible
- Ensuring low-level optimization fairness
- Using the same hardware for all methods
implement all methods in one Codebase
Many good Codebase now for object detection: mmdetection, simpledet, detectron2, etc.
Use the same Codebase to the greatest extent possible.
avoid big tables of historical comparisons
With today’s best practice, the historical results has no meaning to be compared.