[Personal record | Pre-defense for the second year of graduate school]

[ Zhang-Pelvic Test ]

Personal feeling:

1. The explanation must be more skillful and logical; pay attention to the wording (research);
2. The background meaning strengthens the connection with DL. The medical part does not need to be too detailed and is simply listed to give a framework; (Picture Better Text)
3. Research difficulties and challenges, and subsequent research work, should highlight the ① image quality problems of the whole spine (blurring, splicing) ② Compared with spine/skull and other work, the points are scattered and the correlation is difficult to find;
4. List and introduction of the process of data set work;
5. Final implementation - emphasis on intelligent comprehensive reporting;
6. The experiment must have a working foundation; the loss introduction should be more detailed;

Guo-Summary:

1. Background meaning: The explanation is irrelevant, unfocused and complicated; the technical route is explained hastily, not confidently and in depth;
2. The current situation at home and abroad is too simple, which makes people feel that you have read less; you should tell what their shortcomings are and lead to my research work;
3. Difficulties: ① If placed in the whole spine, the film effect will be spliced ​​and blurred; ② Compared with spine detection, the points are scattered and the correlation is difficult to find;
4. Data set work: Annotation and selection of anatomical points can be discussed. For example, DDH is selected at the beginning, and points are added later to list the process.
5. Regarding generating an intelligent comprehensive report, the significance of my research must be highlighted. Because in the end, I didn’t just make a separate data set about the pelvis. It must be placed in the entire pelvic spine. It is an indispensable part. There is a clinical need for this . Others don’t have a data set, so they can’t do it.
6. There must be at least one experiment + idea, and the loss must be clearly stated.
7. The local and global aspects must also be explained clearly. For example, the point detection of the femoral head is Local, and adding global constraints is Global.
8. Reporting is not just a report, the future work will also be writing materials and making reports.

Money:

1. The background and meaning are more about the pelvis and medical significance, and people don’t feel very clear after listening to it (maybe because I find it easy to study by myself, but it is not clear to people who don’t understand it) ; it talks a lot However, the research question was still not raised; moreover, it took 5/6 minutes to talk about the background at the beginning, and the technical route was too fast and was not explained clearly.
2. The difference between the spine and the pelvis is not prominent: the detection of each pyramid point of the spine is relatively concentrated, but the points of the pelvis are relatively scattered;
3. Current research status at home and abroad: The explanation is not sufficient and too simple, which makes people feel that they have not read enough literature.
4. Wording: ① Topic - -> Research ②

Luo:

1. The overall feeling is that I have hastily read some literature, and then did not experiment, and I feel that there is no content;
2. There is relatively little talk about the challenges of the method ; what about the key technologies and difficulties?
3. Experiment: There is only one framework and the preliminary work is seriously insufficient.
4. There is an introduction to experts' prior knowledge, but the explanation of technical implementation is not sufficient and is not reflected.

Zou/Gao:

1. There are many introductions about medicine in the background, but the relevance in deep learning is not prominent; the medical relevance can be mentioned, but it is only used as your research background; your explanation gives people the feeling that everything is meaningful, But they are all mixed together;
2. The content related to the spine can be highlighted: the points of the spine are relatively closely related, while the points of the pelvis are relatively scattered;
3. The process of determining and labeling the data set can be highlighted. Manual labeling is difficult, so machine detection is difficult. You have to analyze where the difficulty lies;

There is very little existing work on the pelvis, and the datasets are not publicly available, and I want to generate a comprehensive report on the whole spine, using the whole spine. (less others, my job)

(Be flexible, have enough confidence, and interact with the instructor; highlight your own work; talk less about details to free up time)

(You can’t say everything is not yours, everything is yours; you must highlight what is yours and what is not yours.)

[ Zhou-Spine Point Detection ]

In the deep learning method, the specific process is introduced

feature fusion

local and global

(Technical route-detailed)

Money:

1. There are many words, too many notes, top-heavy, no focus, and you need to look up at the screen ; fluent but little information (a picture says a lot, but there is no change and contrast in the picture, and the key points cannot be grasped - a static picture tells the story a lot of things);
2. Among the 4 points of background and research route, the technical route talks a lot but only hears a framework;
3. Methods, the methods feel like they are just a summary, and they do not address this problem; a lot of difficulties are discussed, and how to solve the difficulties is not highlighted or innovative in how to solve them;
4. The existing work foundation and experimental design should also be reflected;

Luo:

1. The background and difficulties are explained very well and fully, but the methods are not related to the previous ones; there are too many parts in each part and need to be streamlined and summarized to a certain extent;

open:

1. The previous input was a single plane, but the method only reflected that the input was a double plane. How do you use this double plane to highlight your fusion. Secondly, the effect will be different depending on the location of feature fusion. Why do you put it here? You give people the impression that just by adding a biplane data set, others can do it. So highlight your innovations: 1. Bi-plane feature fusion 2. Local and global features. (Chen Ming has two options.) You can think of various options in the question opening and think more divergently.
2. The explanation is very clear, but the pictures in the ppt are not enough.

high:

1. You have said a lot, but you only have one meaning to express. You can sum it up and give an overview first and then introduce it, otherwise people will not be able to follow;
2. The innovation points should correspond to the difficulties you raise;
3. There are several points in the technical route that can be discussed, which give people the impression that they are just some operations; the meaning of the technical route and how you operate it;

Zou:

1. Technical route: Methods and routes have been talked about a lot, but the correspondence between text and pictures;

Guo:

1. There are too many words in the shortcomings, you can add some pictures; the shortcomings should be matched with innovation [mainly around single shortcomings and double advantages]: ① Occlusion and blurring of single view (one side is blocked and the other side is not blocked) ②; Various method models and finally generate a software;
2. Each group of questions lasts 8 minutes; one is to refine the content, do not go into unnecessary details, and save time to talk about other things.
3. Content and technical route - no need to go into details: what method to use, what functions this method has, and what problems it can solve. The teacher is more concerned about whether your method is effective, whether it can be expanded on existing work, and its significance, etc.
4. The more interactions you have with your mentor, the faster your progress will be.
5. Experiment: if there is one, there should be one. If not, at least there should be a design. Experiment with offset.

[ Guo-Algorithm ]

Money:

1. The report is louder, with a bit of voice inflection and rhythm;
2. Mainly focused on the research content, there are many, and I feel that many of them are lumped together; I feel that it is not thorough enough and superficial;
3. Do not just list articles on the current research status at home and abroad, as the research difficulties are not reflected;
4. The expression is too colloquial, and the "next" point is separated from each other; it should be about what problem should be solved;

open:

1. Gravity algorithm, what others have done, what improvements you have made, and how your improvements are different from others' improvements; it will be more clear if you have a comparison;
2. For teachers who don’t understand your work, they don’t know what exactly you do and what you plan to do.

Luo:

1. The experimental table is not very easy to understand; the meaning of w/t/l; the completed experimental results, the previous one has a great influence on the next one;

Guo:

1. Problem description: When introducing the background meaning, it is a bit complicated and unclear; when placing the gravity search algorithm in image segmentation/threshold segmentation, you must first explain your research background clearly.
2. The teacher will not let you talk about the algorithm process in the question opening;
3. K-best meaning
4. Nowadays, it is generally faster to write questions. Your progress and results.

[ Hui-Video Detection ]

Money:

1. The explanation is not fluent, there are many pauses, and the series is not coherent enough;
2. After talking about the research background, the research motivation (especially regarding the data set) does not need to be introduced too much;
3. The current status of research at home and abroad: we only need to talk about the shortcomings of different supervision methods;
4. The technical route emphasizes solving problems that others have not solved. The focus is on parts 2 and 4, which can be combined when explaining.

open:

1. Highlight detection-->Highlight detection;

Guo:

1. Express coherently and fluently, but will be interrupted by impatient teachers;
2. The motivation part needs to be emphasized again; it is what the overall effect of the method you add is;

Luo:

1. Hasty: The animation is uncoordinated (the P5 ppt data set needs to be synchronized with the picture reality. Pictures are more eye-catching than text, so the teacher has no patience; and the meaning of this ppt is to highlight the modality of this video compared to the original one) Increased richness: complex videos, etc. highlight the difficulty of research)
2. The main content of the research: from supervised to unsupervised based on multi-modality: these two works should be continuously iterated and advanced, and the same should be true for future work;
3. Technical route: You didn’t do it without supervision, but after listing and explaining it, you feel like you have done it;
4. All presentations should highlight key and non-key points.

[ Feng-Human Body Posture Estimation ]

Lightweight HRNet

Money:

1. What aspects should you explain clearly, the research focus of the defending teacher, and your explanation strategy;
2. There are too many modal particles in the explanation (um, ah); causing the teacher’s explanation fatigue; the content should be smoothed out, like communication;
3. Current research status at home and abroad: The explanation is not sufficient and too simple, which makes people feel that they have not read enough literature. And without clearly explaining their shortcomings, they released the research content, which felt out of touch;
4. The technical route explains the models, loss functions, etc. in a very broad and not profound manner. For example, it only gives an MSE formula; you need to highlight your innovation points in the model;
5. If you clearly explain the research content and the reasons for the introduction of the module model, you don’t need to explain so much about the subsequent work;
6. There is no need to go into too much detail about what follows.

Guo:

1. You must be proficient in playing ppt and be fully prepared;
2. The content of the explanation should be refined: for example, if the formula is not distinctive, do not explain it in detail; for the experimental part, do not explain it in too much detail. The teachers who propose the topic only focus on: whether your idea is innovative, useful, and whether it can be realized. Don’t go into too many details. They are not reviewers;
3. The lightweight processing was not reflected in the experimental process.

open:

1. Explain how enthusiasm respects the audience (attracting and highlighting);
2. Ppt layout: position, blank position, font size; ppt is placed very slowly and takes up time;
3. Disadvantages: Use specific pictures to visualize; explain the problem clearly; related work - have you introduced anything related to you? eg lightweighting: why it is necessary to lightweight, what work is done, how they are done, and what are the shortcomings after doing this; the feeling here is that the related work introduced is not related to your work;
4. The research content is just a general title, it doesn’t need to be too detailed. All work must be done with a reason, not just reading a paragraph and telling it to others;

Guo:

1. Research content: Just use the title and don’t go into a paragraph; this kind of routine mistake will be interrupted directly;
2. Bi-weekly report: two days, as detailed as possible;