This is Sudhini from South Bay, California and welcome to my channel where we talk about everything AI.

我是來自加利福尼亞州南灣的 Sudhini，歡迎來到我的頻道，在這裡我們談論人工智能的一切。

And today we are going to be covering a new annotation tool which is scalable and super useful for video images.

今天，我們將介紹一種新的註釋工具，它具有可擴展性，對視頻影像超級有用。

So if you have videos of lung CT, MRIs or even if you have videos of outdoor images or indoor images that you need to be annotated, then this is the software for you.

是以，如果您有肺部 CT、核磁共振成像的視頻，甚至如果您有需要註釋的室外影像或室內影像的視頻，那麼這款軟件就是您的最佳選擇。

The software is called Scalable and I will be walking you through all of the steps right from installation to the requirements and how to annotate stand-alone images, biomedical images to video tracks which is autonomous video tracks.

該軟件名為 "Scalable"，我將指導大家完成從安裝到要求的所有步驟，以及如何註釋獨立影像、生物醫學影像到視頻軌跡（即自主視頻軌跡）。

So if you are an annotator who wants to gain some extra bucks doing annotations at home for some companies or if you're a startup or even an engineer who wants to make sure that you have some group of data that you want to quickly get annotated but you want to ensure quality in this whole process, then this video is meant only for you.

是以，如果你是一名註釋員，想在家裡為一些公司做註釋賺點外快，或者你是一家初創公司，甚至是一名工程師，想確保自己有一組數據想快速得到註釋，但又想確保整個過程的品質，那麼本視頻就是為你量身打造的。

So keep watching and let's get straight to it.

請繼續收看，讓我們直奔主題。

So this is the software that we are going to be reviewing today.

這就是我們今天要介紹的軟件。

It's called Scalable and it's by Berkeley's Deep Drive, so BDD.

它的名字叫 Scalable，由伯克利的 Deep Drive（即 BDD）開發。

All of the documentation that we are going to be following is in this particular guide and I'm going to walk you through each and every one of these steps specifically for my Windows-based system because these directions are mainly for a Linux system.

我們要遵循的所有文檔都在這一特定指南中，我將指導你完成每一個步驟，特別是針對我的 Windows 系統，因為這些指導主要是針對 Linux 系統的。

I will show you the changes that need to be done in order to run it on a Windows system.

我將向您展示要在 Windows 系統上運行它所需做的更改。

So far, I have done another review on another annotator which is called LabelMe and I'm going to be putting the link right up here.

到目前為止，我已經對另一款名為 LabelMe 的註釋器進行了評測，我將把鏈接放在這裡。

So I wanted to show you what the differences are between LabelMe and Scalable before I got to the review.

是以，我想在評論之前向大家展示一下 LabelMe 和 Scalable 之間的區別。

So LabelMe, as you have seen before, it supports single-user annotation and so it will work for single-user situations.

是以，正如你之前看到的，LabelMe 支持單用戶註釋，是以它適用於單用戶情況。

Then it also supports JSON to image label conversion.

此外，它還支持 JSON 到影像標籤的轉換。

So if you have JSON format data, you can also convert that into images, which is typically required for semantic segmentation or unit sort of algorithms.

是以，如果您有 JSON 格式的數據，也可以將其轉換為影像，這通常是語義分割或單元算法所需要的。

You can quickly generate images out of that.

您可以從中快速生成影像。

The installation requirements are pretty on the low side, so it's very easy to install for a standalone computer.

它的安裝要求很低，是以很容易安裝到獨立電腦上。

It is definitely very useful for small-scale projects.

對於小型項目來說，它無疑非常有用。

I have met a lot of developers who are essentially using this for their day-to-day project works.

我遇到過很多開發人員，他們的日常項目工作基本上都是用它來完成的。

And it is preferable for single images.

對於單幅影像來說，這種方法更為可取。

So single images means if you have snapshots of maybe your retinal images, if you have snapshots from different patients, then this is very preferable for those sort of situations.

是以，單張影像意味著如果你有視網膜影像的快照，如果你有不同病人的快照，那麼在這種情況下，單張影像是非常可取的。

Now let's look at what Scalable can support.

現在讓我們看看 Scalable 可以支持哪些功能。

Scalable now uses multi-user annotation.

Scalable 現在使用多用戶註釋。

So if you have, let's say, a huge task, so 900 or 1,000 samples, and all of them belong to particular videos, what it will do is it will automatically create tasks for multi-users, so 3, 4, 5 users, and then each and every one of the vendors will have their own link to go and annotate.

是以，如果你有一個龐大的任務，比如 900 或 1000 個樣本，而且所有樣本都屬於特定的視頻，那麼它就會自動為多用戶創建任務，比如 3、4、5 個用戶，然後每個供應商都會有自己的鏈接去註釋。

So now you can actually scale your work from one annotator to multi-annotators.

是以，現在您可以將工作從一個註釋者擴展到多個註釋者。

You do not require, so in this case, the outcome is actually the output is in JSON format.

您沒有要求，是以在這種情況下，輸出結果實際上是 JSON 格式。

So you cannot really generate images out of that.

是以，你無法真正從中生成影像。

So in order for you to generate, again, images out of the JSON, you will again have to go back to LabelMe.

是以，為了再次從 JSON 生成影像，您必須再次回到 LabelMe。

The third and the most key change in this case, it's actually that Scalable has a Docker container already created.

第三個也是最關鍵的一個變化是，Scalable 已經創建了一個 Docker 容器。

So it has a Docker image that you are going to be pulling from Docker Hub, and that is the one that is going to be running.

是以，它有一個你將從 Docker Hub 提取的 Docker 鏡像，這就是將要運行的鏡像。

So essentially, you do not require any installation requirements on your system.

是以，您的系統基本上不需要任何安裝要求。

If you have Conda, if you have the Anaconda environment, and if you have Docker set up, it will do everything else.

如果你有 Conda，如果你有 Anaconda 環境，如果你設置了 Docker，它就能完成其他一切。

And this is the very latest trend in which most of the applications are going towards.

這也是大多數應用程序的最新趨勢。

So you don't really have a list of requirements that you need to import or all the functions or the libraries you need to import.

是以，你並沒有需要導入的需求列表，也沒有需要導入的所有功能或庫。

You just get the Docker image down on your computer and you just run that.

你只需在電腦上下載 Docker 鏡像，然後運行即可。

So it's very useful and it's very easy to distribute once your Docker container, all the requirements are satisfied.

是以，一旦你的 Docker 容器滿足了所有要求，它就會非常有用，而且非常容易分發。

It is very easy for scale up, like I mentioned.

就像我提到的，它很容易擴大規模。

So once you know how to run your Docker container, you can ask four or five people to do the same process.

是以，一旦你知道如何運行 Docker 容器，你就可以讓四五個人做同樣的事情。

And it is very easy to replicate.

而且很容易複製。

And this is highly preferable for video tracks.

對於視頻軌道來說，這是非常可取的。

So if you're doing video annotation where you have the same objects being followed across different image frames, then that's where Scalable is the most useful.

是以，如果您在進行視頻註釋時，需要在不同的影像幀中跟蹤相同的對象，那麼這就是 Scalable 最有用的地方。

So let's get straight to the installation and the annotation processes now.

現在，讓我們直接進入安裝和註釋過程。

All right.

好的

So in order to install this particular annotation software, which is called Scalable.ai, there are a few things that we have to make sure that your system has first.

是以，要安裝這款名為 Scalable.ai 的註釋軟件，我們首先要確保你的系統具備一些條件。

First off, you need Docker.

首先，你需要 Docker。

And if you're using a Windows system, which I am, I have Anaconda on top of that.

如果你使用的是 Windows 系統，我的系統上就有 Anaconda。

So this would be the command in order to install Docker.

是以，這就是安裝 Docker 的命令。

Now, I will be enlisting all of these commands in the description box below so that you can apply it for your own case.

現在，我將在下面的描述框中列出所有這些命令，以便你可以根據自己的情況加以應用。

So, you know, first of all, let's say that the Docker has been installed.

所以，首先，我們假設 Docker 已經安裝完畢。

The next thing you need a Docker desktop for your Windows, especially if you have a Windows system.

接下來，你需要一個適用於 Windows 的 Docker 桌面，尤其是如果你的系統是 Windows 的話。

So this will take you to this particular page.

是以，這將帶您進入這個特定頁面。

You install the executable.

安裝可執行文件。

And once the executable is installed, especially if you have a Windows machine, it will ask you to also enable the WSL backend, which will ensure that you can run not just in Windows, but also Linux containers.

一旦安裝了可執行文件，尤其是如果你使用的是 Windows 機器，它還會要求你啟用 WSL 後端，這將確保你不僅能在 Windows 中運行，還能在 Linux 容器中運行。

So this is very important for you to do.

是以，這對你來說非常重要。

Once all of this is done, you should then be able to launch your Docker desktop.

完成所有這些工作後，你就可以啟動 Docker 桌面了。

And in the Docker desktop, you can actually go in and generate your Docker login, which will then give you access to all the different kinds of, you know, Dockers that are available to you.

在 Docker 桌面上，你可以進入並生成你的 Docker 登錄名，這樣你就可以訪問所有不同類型的 Docker。

So you have to make sure all of this is enabled first for you to, you know, access this particular annotator, which is which has been pushed as a Docker container.

是以，你必須先確保所有這些都已啟用，才能訪問這個已作為 Docker 容器推送的特定註釋器。

So now that Docker has been installed, let's go on to the next section.

現在 Docker 已經安裝完畢，讓我們進入下一部分。

You start with the GitHub repo.

從 GitHub 倉庫開始。

So this is a GitHub repo that you unpack in the location of your choice.

是以，這是一個 GitHub repo，你可以將其解壓到你選擇的位置。

Now, notice inside this particular folder called scripts.

現在，請注意這個名為腳本的文件夾。

So these are the these are the shell scripts that you need to run in order to ensure that your local environment is exactly what this app requires for it to run.

是以，您需要運行這些 shell 腳本，以確保您的在地環境完全符合此應用程序的運行要求。

Now, the first thing that you that this particular thing is going to do is it's going to set up a local directory.

現在，這個特殊軟件要做的第一件事就是建立在地目錄。

So what this particular code does is it creates this folder called local data.

是以，這段代碼的作用就是創建名為在地數據的文件夾。

And this local data is sort of from where this particular app is going to, you know, access images if you require.

這個在地數據就是這個特定應用程序要訪問影像的地方，如果你需要的話。

So let's say that I have this new batch of images that I need to annotate.

比方說，我有一批新的圖片需要註釋。

I am going to go ahead and put them inside this this local data items.

我要把它們放在在地數據項中。

You know, CT.

你知道，CT。

These are the CT images that I want annotated.

這些是我想要標註的 CT 影像。

So what I do is I go ahead and I put them inside this items folder.

所以我的做法是，把它們放在這個項目文件夾裡。

As soon as I do that, what it does, what the app does is it generates a fake local path corresponding to each and every one of these images.

一旦我這樣做了，應用程序就會為每張圖片生成一個假的在地路徑。

So the app can only access images through URL.

是以，該程序只能通過 URL 訪問影像。

So if you have images that are available sitting in S3 buckets or GS buckets, then you can easily annotate them from there directly using the URL.

是以，如果您的圖片存在 S3 文件桶或 GS 文件桶中，那麼您可以直接使用 URL 從那裡輕鬆地對其進行註釋。

All right.

好的

So once that is done, the you know, the one.

所以，一旦完成，你知道的，那個人。

So once you know you have your batch script and you have your local directory set up, then you need to do a docker pull.

所以，一旦你有了批處理腳本，在地目錄也設置好了，就需要進行 docker 拉取。

So let me go to my Anaconda PowerShell.

所以，讓我來看看我的 Anaconda PowerShell。

And here I am going to be.

而我就在這裡。

First of all, let me activate the virtual environment that where I have Docker.

首先，讓我激活裝有 Docker 的虛擬環境。

Now I need to go to the path where everything is stored.

現在，我需要進入存儲所有內容的路徑。

Right.

對

So documents, annotations, scalable master, scalable and that's it.

是以，文檔、註釋、可擴展的主文件、可擴展的主文件，僅此而已。

So I am currently at the location.

所以，我現在就在這個地方。

So now let me check for Docker images.

現在讓我檢查一下 Docker 映像。

So I have already done a Docker pull.

是以，我已經進行了一次 Docker 拉取。

You see, this is the scalable slash WWW and you see it's 5.12 gigabyte.

你看，這就是可擴展的斜線 WWW，你看它有 5.12 千兆字節。

So make sure you have that much space in order for it to run.

是以，請確保您有足夠的空間來運行它。

The other thing that needs to happen in order for this Docker image to run is, again, there are a few there are a few things that you need to pass to it.

要運行這個 Docker 鏡像，還需要做另一件事，那就是向它傳遞一些資訊。

One of them is a config file.

其中一個是配置文件。

It's the config.yaml.

這是 config.yaml。

So once this is there, I just run this command.

所以，一旦有了這個，我就運行這條命令。

And it is now ready to accept accept any commands.

現在，它已經準備好接受任何指令。

Right.

對

So let's say that I am going to go to this particular page.

比方說，我要訪問這個特定頁面。

And first off, we start by creating, creating projects.

首先，我們從創建開始，創建項目。

Right.

對

So you see, I've already had two projects.

所以你看，我已經有兩個項目了。

But for this case, let me start a fresh one.

但對於這個案例，讓我重新開始。

So CT images.

所以 CT 影像。

Right.

對

And in this case, these are all images labeled type.

在這種情況下，這些都是標有類型的影像。

I'm going to make the polygons.

我要製作多邊形。

It's 2D segmentation.

這是二維分割。

Now, there are three different attribute files that need to be shared.

現在，有三個不同的屬性文件需要共享。

These are the three files that I'll be sharing.

這就是我要分享的三個文件。

First off, if you see, it is it is going to be generating these HTTP links.

首先，如果你看到，它會生成這些 HTTP 鏈接。

These are the fake path links that I talked to you about.

這些就是我跟你說過的假路徑鏈接。

But if I just go to a particular URL and if I just paste this, you'll see that you can now access this image.

但如果我轉到一個特定的 URL，然後粘貼這個，你就會看到現在可以訪問這張圖片了。

So if again, like I mentioned, if you have any image in an S3 bucket, you can just call the S3 bucket path here.

所以，就像我提到的，如果你在 S3 存儲桶中有任何影像，你可以在這裡調用 S3 存儲桶路徑。

And this video name is going to make sure that it belongs to a separate video track.

這個視頻名稱將確保它屬於一個單獨的視頻軌道。

So this is the important image list that you need to pass.

是以，這是您需要通過的重要影像列表。

Then it's the categories that you want to annotate.

然後就是你要註釋的類別。

So it's, you know, ground glass opacity.

所以，你知道的，這就是研磨玻璃的不透明性。

And then there is lung.

還有肺。

That is what I need to annotate.

這就是我需要註釋的內容。

And then the segmentation attributes.

然後是細分屬性。

In this case, I'm calling it if it's blurry or if it's truncated or if there is writing on the image scans, then, you know, behave separately.

在這種情況下，如果影像模糊或被截斷，或者影像掃描上有文字，我就會調用它，你知道，要單獨處理。

So item list.

那麼項目清單。

In this case, it's the image list categories.

在這種情況下，就是影像列表類別。

I already have the categories and attributes.

我已經有了類別和屬性。

Again, segmentation attributes is not a super important file.

同樣，分段屬性並不是一個超級重要的文件。

And now if I say task size, I'll say give me 10 images per task.

現在，如果我說任務大小，我會說每個任務給我 10 張圖片。

And when you say, you know, dashboard, it will take you to the dashboard.

當你說 "儀表盤 "時，它會帶你進入儀表盤。

Right.

對

So here you will see these are the task links.

在這裡，你會看到這些任務鏈接。

So whenever, you know, you are you're doing a particular set of jobs, this will take you to the task.

是以，無論何時，只要你正在做一組特定的工作，它就會帶你完成任務。

So this comes as the first task.

是以，這是第一項任務。

So here you can essentially run each and every one of the 10 images.

是以，在這裡你基本上可以運行 10 幅影像中的每一幅。

Like I mentioned, and see if they are, you know, what you what you need or not.

就像我提到的，看看它們是否符合你的需求。

So let's say that.

那就這麼說吧。

Let's start annotating.

讓我們開始註釋吧。

Once all of this is done, in order for you to now download in CT, you now have this option to download labels.

完成所有這些操作後，您就可以在 CT 中下載標籤了。

If you do this, this is actually going to tell you that in this particular in image 0, these were all the vertices corresponding to, you know, Lung and Activity.

如果你這樣做，它實際上會告訴你，在影像 0 中，這些頂點都與 "肺 "和 "活動 "相對應。

So this could be, you know, again, it is always in the JSON format.

是以，這可能是，你知道，它始終是 JSON 格式。

You are not going to get an image format.

您不會獲得圖像格式。

So you will have to use something else in order to, you know, compute from JSON to your images.

是以，您必須使用其他工具才能將 JSON 計算為影像。

But this is only going to give you because, you know, this particular software is made compatible with autonomous drive sort of scenarios.

但這只是因為，你知道，這個特殊的軟件是與自動駕駛類場景兼容的。

So that's the reason why it will only give you outcome as JSON files.

這就是為什麼它只能以 JSON 文件的形式提供結果。

So now let me do a new one.

所以，現在讓我來做一個新的。

And this new project I'm going to call video bounding box.

這個新項目我稱之為視頻邊界框。

And in this case, I'm going to call it video tracking labeling by bounding boxes.

在這種情況下，我把它稱為 "通過邊界框進行視頻跟蹤標註"。

In this case, the examples.

在這種情況下，舉例說明。

So the so the attributes and everything.

所以，所以的屬性和一切。

It's already an image list which is present.

這已經是一個存在的影像列表。

Again, these images correspond to autonomous drive situation.

同樣，這些影像與自主駕駛情況相對應。

So image list is already there.

是以，影像列表已經存在。

Categories.

類別

I'm just going to call categories.

我只是要調用類別。

Attributes is going to call B box attributes and say go to dashboard.

屬性 "將調用 "B 框屬性"，然後說 "轉到儀表板"。

So as soon as I go to the dashboard, you see the new one is actually created.

是以，只要我進入儀表盤，你就會看到新的儀表盤已經創建。

So now every single one of them will have.

所以，現在他們每個人都會有。

So it has around 23 images per track.

是以，每個軌道大約有 23 幅影像。

So you can literally look at every single one of the images.

是以，您可以查看每一張圖片。

Now, let's say that for this particular exercise, what I wanted to do is I wanted to track pedestrians.

比方說，在這個特定的練習中，我想做的是追蹤行人。

Pedestrians are my regions of interest.

行人是我感興趣的區域。

And I'm going to be creating bounding boxes.

我要創建邊界框。

So.

那麼

So.

那麼

So.

那麼

So.

那麼

If I run this.

如果我運行這個

See up till whenever you annotated it.

直到你批註為止。

Now you can actually see the people being followed by the same color.

現在你可以看到被相同顏色跟蹤的人了。

So you are not going to get a different color bounding box every single time.

是以，你不會每次都得到不同顏色的邊界框。

But you can actually link these pedestrians based off of their movement.

但實際上，你可以根據行人的動作將他們聯繫起來。

So now if this allows for you to generate the mods or any place where you actually need a specific object ID, like these two pedestrians, everyone will have a unique object ID.

是以，如果這樣就可以生成 MOD 或任何需要特定對象 ID 的地方，比如這兩個行人，那麼每個人都會有一個唯一的對象 ID。

You will now be able to run things like that.

現在您就可以運行這樣的功能了。

So if you have aerial view videos or if you have street view videos, then this particular software is super useful.

是以，如果您有鳥瞰視頻或街景視頻，那麼這款特殊的軟件就超級有用。

So this is how, again, once you're done, you just go and hit submit.

就是這樣，一旦你完成了，就點擊提交。

And once it is submitted, you can actually go back and download it.

一旦提交，您就可以返回並下載。

Go to the create page.

轉到創建頁面。

These were the video box and I can just say download.