It all starts with respect. I had last week a recruiter who told me how fantastic my profile is, that he had a mission for me and wanted to telephone. So I reply to his email and offer some time slots for a telephone call. After that, complete silence. I assume that he has found his candidate or that he has shifted his attention to other job descriptions. But to me this in incredibly impolite to not come back and at least give an update like "Hello, thank you for your interest but for this mission we have found another candidate". It actually happens quite regularly and confirms for me that candidates are treated as commodity goods.
Commencez à lire des livres, des BD, et des articles en anglais sur un sujet qui vous intéresse, pour pas avoir des problèmes de motivation. Vous pourrez ainsi construire votre vocabulaire. J'ai fait la même chose pour l'anglais et le français et cela m'a vraiment aidé.
At my company we neither use Yolo nor Roboflow. They're too expensive in a production setting (both from a compute standpoint and a monetary cost standpoint).
In terms of performance, most people think about measures as accuracy, precision, recall, F-means. But one should equally think about the architecture (and energy) that is needed to get to this kind of performance. ViT will need GPU's in most use cases, whereas many CNN's can be trained using CPU's. ViT's are also more prone to overfitting than CNN's when one is using relatively small datasets.
Did you try to train the model for a single epoch? If you know the time for this, you may be able to estimate it for 50 or 100 epochs or any other number.
Of course, it is not ideal but at least it would give you an indication. Most of my own object detection training schemes reach optimal performance on the validation data in less than 100 epochs, and seeing the loss curve after 10-20 epochs you have a good intuition how many epochs are still needed to reach a plateau. Certainly, this may be different in the situation where you have a high number of classes like OP.
What is probably most interesting and difficult at the same time is that CV is a branch that has pretty well advanced the last years, indeed pushed by the developments in deep learning technology. The consequence is that you need to be able to make specialist stuff. Not only in Python but preferably also for production in C++ / Rust since speed and memory are also issues for embedded devices.
It depends on the context of the company and your level of experience in a similar kind of job. If you do not have a mentor, the key is probably finding the right people around you who can help you with single tasks, if needed. This is of course not the same, but at least it will help you speed-up your work and building up a network.
I was wondering in what kind of context or environment are you going to use the computer, what kind of images will your cameras need to register? Are there are any space limitations (restrictions because of small room) or complications for installing the infrastructure. Any other special conditions, such as humidity or high/low temperatures?
Kaggle is a good investment to learn and practice about several topics in ML, such as feature engineering, cross validation, algorithms. It is also fun and can be quite addictive I tell from own experience ;-). However, it is in most cases quite far from real-world ML, where you have to collect and annotate your own data as well as put the model into production. This all takes much more time than the algorithm tweaking that one is doing at Kaggle. And often getting 1% more performance from your model is a serious exercise in Kaggle but in real-world ... a waste of time ;-).
Thank you for sharing! About writing a short email (the approach in 1.): You would say directly that you are interested in the company if there is a job available or would you write an email to get into contact to extend your network? I am just wondering what is the best way to do this and indeed keep it short. If you want to share an example, please feel free!
Wow, impressive result! Small detail: it seems you put Luxembourg at the south-east corner of the Netherlands (province is called Limburg) while it should be at the south-east corner of Belgium.
It appears strange to me that the European map shows the absolute cases per country. Since the total population differs a lot between countries, it might be better to display number / 100,000 habitants.
Yeah, you need your training data to have variations in lighting conditions and car view angles, and probably cameras too. I used to work on one of the globally leading enterprise facial recognition systems. For robust detection and identification we created a "lightscaping room" that a person sits on a chair in the middle, and there are two rigs; one was a bank of lights that can be dimmed, rotated around the subject, composed of a bunch of different types of lighting, and change color; the other rig was a collection of 56 cameras of varying quality from consumer to professional grade that also rotated around the person, independent of the lights rig.
Sounds like a very professional set-up that can only be done in a specialized laboratory. I am curious about the 56 cameras rotating around the person, how many pictures did the cameras take per minute? I guess the system taking the pictures was somehow automated or did you push the button yourself?
As far as I know there are no specific algorithms for this. A way to account for this is using data augmentation as well as include sufficient data examples for training that cover the different lighting conditions and angles.
One downside of Rust is that it misses the libraries and documentation that C++ has acquired over the years. For example, in computer vision it is common to use OpenCV, and this can run in Python and C++. OpenCV does not exist for Rust, you'd need an alternative package.
In essence you a are right but the problem is your word *many*. There will often be packages that are missing or that are in an experimental stage, like OpenCV (
It all starts with respect. I had last week a recruiter who told me how fantastic my profile is, that he had a mission for me and wanted to telephone. So I reply to his email and offer some time slots for a telephone call. After that, complete silence. I assume that he has found his candidate or that he has shifted his attention to other job descriptions. But to me this in incredibly impolite to not come back and at least give an update like "Hello, thank you for your interest but for this mission we have found another candidate". It actually happens quite regularly and confirms for me that candidates are treated as commodity goods.
Commencez à lire des livres, des BD, et des articles en anglais sur un sujet qui vous intéresse, pour pas avoir des problèmes de motivation. Vous pourrez ainsi construire votre vocabulaire. J'ai fait la même chose pour l'anglais et le français et cela m'a vraiment aidé.
At my company we neither use Yolo nor Roboflow. They're too expensive in a production setting (both from a compute standpoint and a monetary cost standpoint).
Could you tell more about the monetary cost standpoint of using Yolo in production setting?
In terms of performance, most people think about measures as accuracy, precision, recall, F-means. But one should equally think about the architecture (and energy) that is needed to get to this kind of performance. ViT will need GPU's in most use cases, whereas many CNN's can be trained using CPU's. ViT's are also more prone to overfitting than CNN's when one is using relatively small datasets.
Did you try to train the model for a single epoch? If you know the time for this, you may be able to estimate it for 50 or 100 epochs or any other number.
Which doesn't really help since you don't know how many epochs will be necessary
Of course, it is not ideal but at least it would give you an indication. Most of my own object detection training schemes reach optimal performance on the validation data in less than 100 epochs, and seeing the loss curve after 10-20 epochs you have a good intuition how many epochs are still needed to reach a plateau. Certainly, this may be different in the situation where you have a high number of classes like OP.
You may want to use Cleanvision
What explains that Switzerland is a bit higher than the neighboring countries?
Everyone has border conflicts with other sovereign nations.
Not an ocean but a sea called the north sea :-) .
Which is a part of the Atlantic Ocean :-) .
And the Atlantic ocean is part of the world. So let's say that "Meanwhile the Dutch has conflicts with the world" ;-) .
What is probably most interesting and difficult at the same time is that CV is a branch that has pretty well advanced the last years, indeed pushed by the developments in deep learning technology. The consequence is that you need to be able to make specialist stuff. Not only in Python but preferably also for production in C++ / Rust since speed and memory are also issues for embedded devices.
It depends on the context of the company and your level of experience in a similar kind of job. If you do not have a mentor, the key is probably finding the right people around you who can help you with single tasks, if needed. This is of course not the same, but at least it will help you speed-up your work and building up a network.
Is this video real or was it AI generated? ;-)
I was wondering in what kind of context or environment are you going to use the computer, what kind of images will your cameras need to register? Are there are any space limitations (restrictions because of small room) or complications for installing the infrastructure. Any other special conditions, such as humidity or high/low temperatures?
Do you recruit freelancers and developers on remote-basis as well? That may help to find candidates.
Kaggle is a good investment to learn and practice about several topics in ML, such as feature engineering, cross validation, algorithms. It is also fun and can be quite addictive I tell from own experience ;-). However, it is in most cases quite far from real-world ML, where you have to collect and annotate your own data as well as put the model into production. This all takes much more time than the algorithm tweaking that one is doing at Kaggle. And often getting 1% more performance from your model is a serious exercise in Kaggle but in real-world ... a waste of time ;-).
How do you apply to 10+ jobs in case you need to write a motivation letter as well? Do you skip the letter?
I use use
Thank you for sharing! About writing a short email (the approach in 1.): You would say directly that you are interested in the company if there is a job available or would you write an email to get into contact to extend your network? I am just wondering what is the best way to do this and indeed keep it short. If you want to share an example, please feel free!
The build system is called Bazel
Anyone who has done it with Cmake? I have tried it but did not manage.
Wow, impressive result! Small detail: it seems you put Luxembourg at the south-east corner of the Netherlands (province is called Limburg) while it should be at the south-east corner of Belgium.
It appears strange to me that the European map shows the absolute cases per country. Since the total population differs a lot between countries, it might be better to display number / 100,000 habitants.
opencv
Keigo Higashino (Japon), si vous aimez les romans policiers. Commencez avec la série Physicien Yukawa (
Je dirais à 11h00.
Yeah, you need your training data to have variations in lighting conditions and car view angles, and probably cameras too. I used to work on one of the globally leading enterprise facial recognition systems. For robust detection and identification we created a "lightscaping room" that a person sits on a chair in the middle, and there are two rigs; one was a bank of lights that can be dimmed, rotated around the subject, composed of a bunch of different types of lighting, and change color; the other rig was a collection of 56 cameras of varying quality from consumer to professional grade that also rotated around the person, independent of the lights rig.
Sounds like a very professional set-up that can only be done in a specialized laboratory. I am curious about the 56 cameras rotating around the person, how many pictures did the cameras take per minute? I guess the system taking the pictures was somehow automated or did you push the button yourself?
As far as I know there are no specific algorithms for this. A way to account for this is using data augmentation as well as include sufficient data examples for training that cover the different lighting conditions and angles.
Iceland.
The Netherlands. Affordable, not too hot, you'll find community of digit nomads in the major cities.
One downside of Rust is that it misses the libraries and documentation that C++ has acquired over the years. For example, in computer vision it is common to use OpenCV, and this can run in Python and C++. OpenCV does not exist for Rust, you'd need an alternative package.
Many C and C++ libraries have Rust bindings on Crates.io, so this isn't necessarily true
In essence you a are right but the problem is your word *many*. There will often be packages that are missing or that are in an experimental stage, like OpenCV (