Today I want to present a noncontroversial topic by a possibly controversial person:
https://pjreddie.com/media/files/papers/YOLOv3.pdf
A while back, J. Redmon and co-authors published their first object detection architecture, YOLO. The second iteration offered a significant improvement in both speed and accuracy. The current version presents itself as an incremental improvement, but there are other opinions about it, namely:
https://mc.ai/yolo3-a-huge-improvement/
Questions:
1. Do you think such great networks, not using capsules or GANs have future?
2. What's your opinion about the article style?
3. What are your ideas of possible applications of YOLO?
I can say that I have heard about YOLO2.0. At the end of May I have an economic exhibition. I will display my products. I have a plan to use the YOLO technology for face detection, then I want to attach faces to Clouds / speech bubbles with the comments, eg: Maybe it's time for a new pen? or: But great fountain pens! I will let you know how it will work. :)
ReplyDeleteGood idea, but note that you'll need to train your own model for face detection (maybe on one PASCAL VOC dataset, where there were body parts - hands, feet and faces - labeled for training).
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteWhy not? It already proven that it can be used to something useful so, in my opinion, this is just a matter of time when we will find more usages for it.
2. What's your opinion about the article style?
The style is fascinating. I never thought that you can publish something in such informal way. Recently I was publishing my own articles and I was fighting for nearly a month to get everything as the conference publisher wanted to. Every coma had to me on proper place, every sentence had to as formal as possible.
3. What are your ideas of possible applications of YOLO?
What can I say? Mobile augmented reality. Why? Because of the frame rate that they are achieving. Even in first version they were able to process images at 45 frames per second. That is awesome result. The whole approach to mobile ar can be changed thanks to that
Good point - AR can benefit greatly from reliable, but moreover, fast object detection.
DeleteAs for style - Redmon is an extreme case, but making things as formal and boring as possible is a camouflage technique for hiding the scarcity of substance.
Not only hiding but also making everything much longer than it should be. Most of the articles can be and should be much shorter, but the formal language makes it impossible
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteOf course, as long as they work properly. As you can see, their solution in terms of error doesn't stand out so much.
2. What's your opinion about the article style?
The style of the article is very interesting and fits well with the model name itself. I have to admit that I haven't seen an article written in the same way before. However, I prefer such an article to be an interesting exception, not a rule.
3. What are your ideas of possible applications of YOLO?
Taking into account the speed of their solution, they can be used wherever we want to quickly identify the situation. If it is accurate enough, it can be used, for example, in autonomous cars and their sensors.
Thank you for replying. As for the error percentage, the COCO has recently tweaked it's metric (the whole mAP vs meanAP question), and it contains absurdly hard images, i.e.: http://cocodataset.org/#explore?id=300950
Delete(first example that caught my attention when going to the 'browse' option for the 'person' class).
1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteYes, they have. I am not sure are they as good as described in an article. It depends from the analyzed collections. We can find some points in an article that shows the problems of this approach. For instance mentioned: problem with small objects that appear in groups, such as flocks of birds. I believe we have to observe this method because it looks very promising. Nevertheless speed of learning is not such important as accuracy in my opinion. Especially in a object recognition - where accuracy in most of cases is crucial.
2. What's your opinion about the article style?
Layout of the article is clean. It highlights most important assumptions, pros and cons. I think it is modern template, but I am not sure is it a good choice if you want to print it.
3. What are your ideas of possible applications of YOLO?
Each solution based on image recognition. For instance car plates. But yolo mostly will be used in areas where patterns evolves dynamically, because there time that is needed to train a model is most important factor.
Thanks for commenting. I'd like to point out, that YOLOs are remarkably fast for inference (model evaluation), but not exactly for training. The training can take, in my experience, from 2 days to 2 weeks. But I don't (yet) know how it stacks up against e.g. Facebook's Detectron (my colleague trains it, so I'll know in a couple of weeks - if you want to know the results, contact me then).
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteThe learning speed is one thing and accuracy is the other. There are fields of knowledge that require speed and those that require accuracy. YOLO is an interesting example of combining accuracy with speed but a lot, and in this case will depend on the patterns used.
2. What's your opinion about the article style?
The style of the article is interesting. I think we all have quite standard, standard styles of publishing. Maybe it's time for revolutions in this matter too. Only we have to be reviewed in order for this transformation to take place.
3. What are your ideas of possible applications of YOLO?
After such a presentation of the problem, we can think about the wide application of YOLO. Each modern system is based in part on even a part of the code that learns. I believe that there are projects that will enable the full use of YOLO. For me personally the biggest problem is speech recognition and here I see the application for YOLO.
Thank you for responding. I guess, when Redmon and people like him becomes a peer reviewer, we'll be able to write a little more like him.
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteYes probably they have future. If there will not occur anything more sufficient they seems to be one of the most used ones. I believe we have to observe this method because it looks very promising. However this is just one of the current trends and we all know that very often the most popular options are used instead of the most useful ones.
2. What's your opinion about the article style?
It is definitely different than those I have seen before. It is rather short, but at the same time very understandable, using graphics and tables makes it clear. I think it is very good, because it sums up the topic without unwanted beating around the bush.
3. What are your ideas of possible applications of YOLO?
Everything connected to image recognition and processing. Maybe in autonomous cars. Also it might be used in areas where patterns evolves dynamically.
Thank you for weighing in on the topic.
DeleteThis comment has been removed by the author.
ReplyDeleteAd1.
ReplyDeleteThe use of such a neural network has an infinite number of applications. I do not know about this type of algorithms, but from what I have read it is used in recognizing images. This algorithm can be used for good purposes, such as searching for missing persons.
Ad2.
The article shall be legible. Characteristics are shown in an accessible way. This Article is addressed to experts. A person who has no experience in this subject will not gain any knowledge by reading it.
Ad3
If I found an implementation algorithm, I would be happy to use it in my research.
There is an implementation by Redmon on his github (darknet), and for Keras (YOLOv2 only) there is YAD2K. If you want to build on Windows, try out AlexeyAB fork, which follows JPReddie closely, possibly adding some own ideas.
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteHmm but YOLO is basically an object detector and Capsule Networks are derivatives from CNN (are classifiers), GAN are also completly different stuff. But yes when it comes to object detection it can be future of this kind of solution.
2. What's your opinion about the article style?
It is readable and easy to understand so yes I can read more of this kind of documents. You know in what journal it was published?
3. What are your ideas of possible applications of YOLO?
In every application that uses detection (I think that Dawid P will put it into VR :D ). But I think if you can use CNN in NLP than this kind of stuff can have a lot of applications. And remember "The street finds its own uses for things." - Wiliam Gibson :D
Ad 2. It was accepted to arXiv, which is now not automatic anymore (needs reputation or a peer recommendation, but not necessarily full peer review). I'm not sure about journals or conferences. Especially the third article could qualify as 'technical report'.
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteIn my opinion object detection and classification models such as YOLO do have future and with further development could become state of the art in some filed or applications.
2. What's your opinion about the article style?
It was nice read. The comparison seems to be pretty informative but on the other hand target audience is rather people who already have some expertise in this field.
3. What are your ideas of possible applications of YOLO?
Object detection or recognition in:
- autonomous cars
- drones
- security and surveillance
Thank you. Ad 3 - you now mentioned surveillance, and this happens to be what my colleagues an I are using it for currently ;)
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteI think that such networks take place in the future.Since there are few alternatives, I think that they have a huge potential. But unfortunately I do not understand much about this topic, and I'm my subjective opinion
2. What's your opinion about the article style?
Excellent article, I liked the style of writing, very readable. Everything is written very clearly and specifically
3. What are your ideas of possible applications of YOLO?
In the processing and manipulation of data. Products associated with image recognition. Perhaps the vehicle sphere.
Thank you for replying.
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteYes, it seems to be so. As some have already said, there is a trend of using this technology. It looks very promising and is used in many modern solutions.
2. What's your opinion about the article style?
I’m surprised! I probably have not seen much in my life but the informal style in the published article is new to me. In my opinion it is unacceptable - maybe I'm just used to formal style... On the other hand I agree with opinion that this article is addressed to experts. Thus, the article combines two realities: a simple language with advanced content.
3. What are your ideas of possible applications of YOLO?
I am considering the possibility of using YOLO where video monitoring is used. This could support the work of security guards or perhaps even replace them. If not for frequent errors in YOLOv3, I would suggest this technology also for the dogcatchers, but from what I have seen, many objects are unfortunately recognized as a dog ;)
About wrong classifications - well, we could address it with a simple classifier behind YOLO. The detection is the hardest part.
Delete1. I don't know what capsules or GAN is, it is not described well in the article
ReplyDelete2. I think it's not bad if you are interested in the topic, and have some backgound knowledge about yolo, sadly i don't have any
3. Like others said i would use this technology in somekind monitoring features, or speed detecteors maybe, to catch registration plates...
Capsules and GANs are more current, 'cool' research topics in Deep Learning.
Delete1. Yes, I think they have future - interestingly, one "deep" network can detect multiple objects at once or multiple variants of the same object. However, in order to learn such a network, it should be trained on hundreds, thousands or n hundreds of thousands of images. It takes a lot of time, but graphics cards are used for this, which greatly speeds up learning. Of course, besides learning, the tedious process is to prepare a set of data, which usually occurs in the form of a picture with rectangles marked on it, covering objects of a given class (eg vehicles) or in the form of cut-out rectangles. It takes a dozen or so hours of tiring work to exclude a few thousand photos by one person, and so many photos are not very much, and if we want to recognize, for example, everyday objects, say 20 different, we have clicks for weeks or months. Fortunately, there are ready databases of images and their sequences.
ReplyDelete2. The style is very helpful enabling to go through technological issues like through adventure book.
3. I always look for using such ideas first of all in medicine. I guess this is the direction we should try to run, so as the technology could serve to humans. I am not specialist, but my intuitions tells me, that we can gather medical practical information and through learning thanks to such real data, we can be closer and closer to solving all treatment problems.
You're right about the tagging effort. After tagging augumentation can help a bit with increasing training data volume, but the augumentation itself is at least two clicks per object (top left corner, bottom right corner). There are a few pieces of software for that, but none are great.
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteYes. I do. Each new implementation is getting better. This shows that the algorithm has potential and I am curious how it will develop.
2. What's your opinion about the article style?
The article is pleasantly read. The style of writing allows for a simple comparison of different versions of the algorithm.
3. What are your ideas of possible applications of YOLO?
I think that the algorithm will be applied wherever the detection speed of objects counts
Thanks for commenting.
DeleteQuestions:
ReplyDelete1. Do you think such great networks, not using capsules or GANs have future?
I think that this is an interesting idea and this networks have future. Usually some modifications help to bring ideas to real life development.
2. What's your opinion about the article style?
The article style is not the standard scientific style. Abstract seems to be rather an lead to sensational story than to scientific research. I think that it is intentional treatment to bring attention but quite bold as some may not take it seriously.
3. What are your ideas of possible applications of YOLO?
There are many possible applications, for example augmented reality application for museums to recognise objects quickly and get the information about it.
Thank you for commenting. I was once involved in a system for a museum, and it probably would have done better if we had YOLO then.
DeleteBefore answers: is it some insider joke to crop the top part of network schema (https://arxiv.org/abs/1506.02640) as a tribute to Krizhevsky, Sutskever and Hinton article on ImageNet ?
ReplyDelete1. Do you think such great networks, not using capsules or GANs have future?
I think the field is too young to definitely cross out some architectures.
2. What's your opinion about the article style?
I like it. It sort of breaks the 4th wall, and still conveys all information it is supposed to. For those who like this style, there is Revd. Manul Laphroaig and his "POC||GTFO" magazine which takes it to the extreme.
3. What are your ideas of possible applications of YOLO?
It's execution is fast and predictible, I guess some mass tagging of images maybe?
I took a look at the "POC||GTFO" journal, and it is way above and beyond what Joe is doing in his articles. But many thanks for pointing me to it anyway.
DeleteCapsule networks are still on the early stage of evolution, the first paper about them was published a few months ago and the concept was presented on a toy dataset (MNIST). I'm not aware of any example where capsule networks outperform regular networks. And as for GANs, I'm not sure if they're particularly useful in object detection, but I've seen some papers about using them to generate a better representation for such a task (small object detection: https://arxiv.org/pdf/1706.05274.pdf). Therefore I think that other networks won't be demoted because of capsule architecture or GANs.
ReplyDeleteAs for the article's style - it's great! It's entertaining to read and at the same time covers all important details. I wouldn't protest of such style was widely accepted considering, of course, that informational layer of the article is intact. Articles like that are easier to remember. Books from "Head first" series are in a similar style for a reason.
Maybe it can be used for fine-grained video tagging or surveillance.
Thanks for the link - super-resolution is always interesting. For people who want to learn something about GANs, this may be one practical example.
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteWhy not? If something is useful, it will be applied and more and more applications for such thing will be found. If something is proved to be pointless, it will be quickly forgotten.
2. What's your opinion about the article style?
The form is really unique and original, nevertheless it's not appealing to me. But that's just me. I think that "traditional" academic language is used for specific purpose and has long history and it's simply good to be challenged from time to time when reading something instead of surrounding ourselved with simplified versions of simplified versions with simplified style.
3. What are your ideas of possible applications of YOLO?
Definitely all areas related to image recognition and processing.
Thanks for your opinion. If I understand correctly, you ascribe intellectual challenge to reading formal 'traditional' articles. This would mean, that legal documents and the like (agreements, policies) have such qualities also.
DeleteThank you for an interesting article. I believe that this area of research has enormous potential, especially the armies that want to use military robots in the future may be interested in this area of research. The style of the article is quite original and it is nice to read. Other applications of YOLO - autonomous transport...
ReplyDeleteThank you for commenting.
DeleteThis comment has been removed by the author.
ReplyDelete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteIt's hard to say. This is far from my field of study.
2. What's your opinion about the article style?
Refreshing? It's hard to imagine that reviewers will actually accept it. In my honest opinion, they won't. Unless author will find an open-minded reviewer because this article is written like an email to a friend or something similar.
3. What are your ideas of possible applications of YOLO?
Maybe a new facial recognition software, we are still far from being accurate. I would like to see a prototype to judge.
Maybe Redmon could use help from an editor, who could translate plain English into 'scientific'.
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteI think it is a matter of time when we find more applications for this, provided the article tells the truth and the network is so good.
2. What's your opinion about the article style?
The style of the article is very interesting. It highlights most important assumptions, pros and cons. The article reads pleasantly.
3. What are your ideas of possible applications of YOLO?
If they solve problems so quickly, they could be used to create augmented reality system.
One way of telling the quality of the solution is benchmarking on a public dataset. However, the problem is multi-criteria, so sometimes solutions with different trade-offs are not directly comparable.
DeleteDo you think such great networks, not using capsules or GANs have future?
ReplyDeleteWhy not! We had a lot of this kind of technologies like Bluetooth, 802.11. Both of them don't have to use GANs but they still have future and - I guess - this one can be the next one - why not! Again.
What's your opinion about the article style?
It's an excellent article. The correct layout of the text shows the article was published in a prestigious magazine. Abstract with a thesis and punch line, diagrams, a chronological structure provides us step by step cross the hard subject. Especially for me because this kinds of articles aren't my favourite.
What are your ideas of possible applications of YOLO?
No idea. As a lawyer, I can say it's too hard for me. I can only assume it made me impressed what I've read about this, but half of the text is out of my mind scope.
Thank you for your kind and light-hearted comment.
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteI believe that the main area of interest is objects detection and classification problems.
2. What's your opinion about the article style?
I enjoyed it, nonetheless, one has to have a knowledge in this field to benefit from this article. It is not a traditional style. Probably it the cause why it was so interesting.
3. What are your ideas of possible applications of YOLO?
Of course autonomous cars and security are the best examples of the recognition activities. Further development of the Augmented Reality I think too.
Thank you for commenting.
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteYes, While I was reading this article an idea apperead in my head about where I can use free neural network darknet and YOLO. Right now I would rather not write about it more specifically, because I must do some deeper research about implementation, specially on Android system, before I could decide, if the idea will be effective.
2. What's your opinion about the article style?
Style of this papper is very informal and extraordinary. At first I think it is very strange, but after reading it all I am sure no one can be bored while reading it. What’s most important, the style of the article is no obstacle for understanding its content.
3. What are your ideas of possible applications of YOLO?
Like I write in my first response, I think I can use darknet and YOLO in my future research. I am also agree with the possible applications of YOLO metioned in the Ewelina Winska’s comment.
I wish you good luck with this research. There is a mini-YOLO port for Android, which you can check out and maybe save some time.
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteI'm not an expert in that field but I agree with your opinion that you've put in the question. If something is great I think that it has future :) I'll go deeper with the YOLO cause it looks very promising.
2. What's your opinion about the article style?
Cool and loose style! It is one of the few papers that after reading you can actually code something.
3. What are your ideas of possible applications of YOLO?
I think that automatic content moderation in the social media could be one of the examples.
Thank you for an interesting idea, which has not been mentioned here yet. And it inspires new ideas - identifying spam images or for example illegal items in online auctions.
Delete1. Do you think such great networks, not using capsules or GANs have future?
ReplyDeleteI think that yes. I don't have idea on which task they could be better done but I'm sure it is. Maybe in future their possibilities would be improved.
2. What's your opinion about the article style?
I am a bit confused, but I have to say it was nice to read. All the information was exhaustive, and yet I did not feel that I was reading a scientific article.
3. What are your ideas of possible applications of YOLO?
They are really fast so in business application which can prognose exchange rates.
Thanks for commenting, albeit I'm hard pressed to imagine the application for exchange rate prediction.
DeleteThank you for your articles that you have shared with us. Hopefully you can give the article a good benefit to us. Eilat Lev Eilat Lev Scarsdale
ReplyDelete