Frequently asked questions.

Getting Started

Matroid is a Computer Vision AI company whose no-code software gives users the power to build, deploy, and maintain detectors, allowing companies to search visual media for people, behavior, objects, defects, and events. Based on the latest in deep-learning and computer vision research, Matroid’s capabilities can improve processes, safety, and operational efficiencies across industries including manufacturing, healthcare, security, agriculture, and transportation.

How does Matroid work?

Matroid is an easy-to-use and intuitive studio for creating and deploying detectors (computer vision models) to search visual media. Deployed in the cloud or on-premises, the technology is camera agnostic and works with any image or video sources. It also leverages on-premises deployments so companies can keep data private and secure with no internet required.

Users can build, test, deploy, and manage detectors to help make sense of visual data, which then provides real-time alerts, analytics, and summaries on events to better understand what happened before, during, and after those events or conditions.

Is Matroid’s technology complicated?

Not at all. Matroid’s easy-to-use point-and-click capabilities mean that you don’t need to be a data scientist to use and understand the platform. And it can be integrated into existing tools and cameras, further simplifying how to deploy and use it. Best of all, its no-code platform allows users to become computer vision experts and deploy detectors in a matter of minutes rather than months.

How do I know if my company could benefit from Matroid’s technology?

Ask yourself the following questions:

Are the challenges my company is facing visible by some form of camera technology?
In what aspects of my company’s operations can image or video detections bring value?
What insights do I want to gain from computer vision AI?
Can this application span across multiple operations?
Are there areas within my company where we can improve operational performance, safety, and the bottom line by simply better understanding our operational efficiencies?
If I had an automated pair of eyes, what would I want to visually monitor in my operations?

What’s the difference between Computer Vision and Machine Vision?

The main difference between machine vision and computer vision is the level of intelligence and autonomy exhibited by the systems. Machine vision is a more limited form of visual processing that is traditionally used in industrial automation and is typically based on rules-based tools. Computer vision is a significantly more sophisticated and complex form of visual processing that is based on deep-learning and used in a wide range of applications to enable computers to "see" the world like humans do.

Machine vision systems can recognize basic parts or feature presence/absence. Computer vision systems, on the other hand, can identify and classify things about visual data in a much more advanced way by using deep-learning techniques to identify objects, recognize patterns, and make predictions based on visual data. Computer vision also provides analytics on the defect classifications. This makes computer vision systems much more versatile and adaptable than their machine vision counterparts. For example, machine vision can tell you if a hole is present or absent, but computer vision will classify weld defects (spatter, porosity, burn through, etc.) and precisely locate those defects.

What industries are currently using Computer Vision?

What are some of the perks of having a platform like Matroid in my company?

Matroid provides real-time alerts and feedback for machines, systems, and operations which means you can make important decisions or quick changes that can positively impact operational efficiencies, quality control, regulatory compliances, and workplace safety.

Matroid also automates quality inspections. This feature eliminates subjective decision-making by operators whose attention may deteriorate over time, while deep-learning can improve over time.

It can also safeguards tribal knowledge, as it learns and retains visual inspection requirements in complex production environments - similar to a human QC inspector. This means that regardless of your employee turnover, data is protected and doesn’t leave with any one person.

A key feature is that Matroid makes sharing data easy by creating collaborations between cross-functional teams and allowing key players to access detections in real-time.

How do I sign up for Matroid?

How do I get started with Matroid?

Which devices and browsers does Matroid work best with?

How do I register a live feed on Matroid?

Is there someone at Matroid that is available for speaking opportunities, including keynotes, panels or other presentations?

Are there any case studies or documentation that proves the effectiveness of Computer Vision?

How is Matroid’s Computer Vision technology or processes different or better than competitors or alternatives?

Matroid is the first end-to-end CV platform having the greatest industry experience due to its unique, easy-to-use, and accessible platform that is helping customers easily make sense of their visual data.

Since 2016, our patented technologies and solutions have been providing the most sophisticated point-and-click experience for domain experts to identify what they want to detect and classify from visual media. More specifically, our capabilities allow customers to stream monitor, get real-time alerts and

Creating a detector

How do I make a custom detector (model)?

With Matroid you can create a custom detector that can detect whatever you're interested in.

To create a custom detector, click on "Create a new detector" on the home page. Enter labels for the objects you'd like to detect. For example, this detector will be able to detect tigers and lions:

Next, you'll be taken to the detector creation wizard. If there are public detectors that detect similar things, you'll be shown a list of suggested detectors:

If none of these detectors meets your needs, click on "No, continue building detector" at the bottom. Follow the setup steps in the wizard to create a new detector. Once setup is complete, you'll be taken to the studio, where you'll be able to upload and curate training data before training your new detector.

How do I add labels to a detector?

How do I make a detector that can detect specific faces?

To create a detector that can detect specific faces, create a detector with a label for each person you'd like to detect. In the detector setup wizard, select "facial recognition" as your detector type. Your training images should clearly show the face of the person you are trying to identify. If there are multiple faces in a training image, draw an annotation bounding box around the face of interest in the studio.

How can I annotate training images to improve my detector?

You can improve detector accuracy by telling your detector which sections of an image to focus on during training.

Clicking an image will open the image slider view. In the slider view, you can drag a bounding box around the pertinent part of the image. For example, if you were creating a detector with a "male lion" label, you'd want to draw an annotation box around the male lion:

How can I upload more training images for a detector?

How can I upload images from a live feed for training data?

How do I automatically annotate training examples in the Matroid Studio?

When you are creating a detector in the Matroid Studio, you can quickly annotate your training examples using another detector.

First, select the images you would like to automatically annotate. By default, all images on the current page will be annotated. If you would like to annotate a subset of images on the current page, click and drag a box around those images. If you would like to annotate the images from all pages, check the "All pages" check box.

If you are creating a facial recognition or facial characteristics detector, simply press the "Annotate faces" button:

If you are creating a natural imagery detector or object localizer, first click the "Annotate" button:

At first, a list of common objects to localize will display in the popover. If one of these common objects is what you are trying to annotate, click on it:

If not, enter the object you want to localize, and choose a detector from the list:

Note: this will only work if there is already a localizing detector that can find the objects you are trying to localize. For example, if you were creating a cat breeds detector, you could use Everyday Objects to annotate the cats in your training examples.

After you've selected a detector, click the blue button to annotate the selected images:

The chosen detector will start localizing, and annotation bounding boxes should appear around the relevant objects in your training examples. Auto annotating an image with a detector uses the same number of credits as classifying an image with that detector.

If you are creating a natural imagery detector, your trained detector will not localize objects even if you annotate all of its training examples. To create a detector that can localize arbitrary objects, you must choose "Object localizer" as the detector type.

How can I create a detector from a TensorFlow protocol buffer or SavedModel?

You can import a TensorFlow protocol buffer or SavedModel directly into Matroid to use as a detector. To import a neural network, sign into your account on matroid.com, and click "Create a new detector" on the home page. From there, click the "Import neural network" button.

When importing your neural network, you'll specify a few details like a detector name, type, and the labels. You can either manually enter the labels, or upload a labels text file with one label per line. Additionally, if you're creating an object localizer, you can upload a label indices file, where each line in the file is an integer that corresponds with the index of the label at the same line in your labels file.

After specifying this information and uploading your model file, you'll need to specify a few details about your model's input and output tensors. From there, hit "Create detector", and after a few minutes you should be directed to a new detector for your model.

For a complete list of model requirements, view our model import requirements.

Using a detector

How do I classify images and videos using the API?

Matroid's developer API is designed to make it easy to bring the power of computer vision to your app. From the detector page, click on the “Call API” tab to see examples of how to use the Matroid API.

How do I run a detector on a YouTube video?

My image is very large and is causing a slow upload. Is it safe for me to resize the image before uploading it?

How do I run a detector on a live video stream?

Improving a detector

My detector doesn’t work well on images and videos. What can I do to make it better?

The most common cause of poor detector performance is low quality training data. To improve an existing detector, redo it, paying attention to the following guidelines for training data:

A label's training images should be accurate examples of that label. You should include a variety of training images that show the object you are trying to detect from different angles and in varying lighting conditions.
It may be helpful to have images in your training set that are similar to what your detector is likely to encounter in the field. For example, if you are going to run your detector on an outdoor camera feed, you should include training images with varied lighting conditions simulating daytime, evening, and night conditions. It would also help to add actual images from the outdoor camera as training examples, if available.
When a training image contains extraneous details or objects in addition to the object of interest, you should add annotation bounding boxes around the object of interest. This helps your detector focus on the relevant part of the image during training.
Adding negative examples to your detector's labels can help improve detection quality, especially when your detector is getting confused by examples of objects similar to the object you are trying to detect. For example, if you are building a dog detector, you might want to add images of wolves as negative images for your dog class. Providing feedback for incorrectly classified images will also help improve detector accuracy when you retrain it.
Make sure you have the appropriate number of training images for the detector type you are trying to create. Matroid recommends the following minimum numbers of training images:

I tried all of the things above but my detector quality didn’t improve much. What else can I do?

Are there any categories of objects that are impossible to detect with Matroid detectors?

Publishing a detector

How do I share my detectors with the Matroid community?

You can request to publish any of the Matroid detectors you have created that meet the publication criteria listed below. If the publication request is approved, your detector will appear in Matroid’s list of public detectors.

To publish one of your detectors, go to the “My Detectors” page and click on the detector’s name. Click on the "Publish" button and then fill out the name, description, and categories for your detector. This information will help other Matroid users find your detector.

To make a publication request, your detector must meet the following criteria:

It must have been trained with at least 20 images per label.
It must have a training accuracy of at least 90%. If your detector has lower than 90% accuracy, you will see a warning on the detector page.

The Matroid team will review your detector for quality, uniqueness, and appropriateness. You will be notified by email once the request is approved or denied.

Are there any benefits to publishing my detectors?

Streams

How can I register a stream to record screenshots from? E.g. https://206.176.34.51/mjpg/video.mjpg

From the streams page, click on "Register new stream". Enter the stream information accordingly:

Enter the details for your camera and the live stream should display. To make a test recording from your stream, click on the stream name. Then, click on the "New" button under "Recording" in the sidebar, and select a duration.

To view your recording, select it and click on "View Recording Slideshow".

What kind of streams does Matroid support?

Matroid natively supports a wide variety of streams.

If you are creating a stream from a URL, you can provide a URL for:

An HTTP based stream like MJPEG over HTTP, HLS, MPEG-DASH, JPEG over HTTP
An RTSP stream. E.g. rtsp://mysite.com/video
An RTMP stream. E.g. rtmp://mysite.com/video
A Youtube URL

Note that other types of URLs may work even if they're not officially supported.

You can also stream RTMP video directly to Matroid. When creating your stream, select RTMP from the Stream Source dropdown. Click “Register stream”, and you'll be given the URL to stream to as well as a unique RTMP key for your stream.

Finally, you can create a stream from an IP camera. To do so, select "Other Camera" from the stream source pull down. Select the make and model of your camera, and then enter the necessary IP credentials.

How can I stream live video into Matroid via RTMP(S)?

You can stream live video into Matroid via RTMP or RTMPS by creating a new stream, and choosing RTMP / RTMPS in the Stream Source dropdown. Click “Register stream”, and you'll be given RTMP and RTMPS URLs to stream to as well as a unique RTMP key for your stream.

Note that RTMPS is a secure implementation of RTMP over a TLS / SSL connection. Streaming over RTMPS will only work if it is supported by your streaming application. We recommend using OBS at version 24 or higher.

Computer Vision Kit

What is a computer vision kit, and how do I connect it to Matroid?

Matroid makes computer vision simple. We've built an easy-to-use and intuitive studio for creating and deploying computer vision detectors. From there, you can use these detectors to search visual media for objects, animals, people, even specific faces - no programming required!

Follow the computer vision kit's audio instructions along with this guide to create a stream. By default, we use our own Everyday Objects detector to find people and cars on your new stream. Any time a person or car comes into the camera's field of view, a detection notification will show up on both the navigation bar and the stream's calendar page. It's that easy!

Setup

Power on the camera and remove the plastic film from the lens.
Wait a few seconds for the audio instructions to begin.
Navigate to /camera, then either create an account or login with an existing account.
Register the camera by entering a stream name and your WiFi settings.
Click “Next” to generate a QR code, then scan the code when instructed.
Once the camera has connected to Matroid, you'll be automatically redirected to the new livestream.
Follow along with the stream tour to learn a bit more about the available features.

Will the camera setup be any different for an on-prem machine?

On-prem users should follow the same setup process, with some additional steps:

Open up ports 80, 443, and 1935 by modifying your system's firewall settings. Run the following commands for these three port numbers:
- CentOS
  - sudo firewall-cmd --zone=public --add-port=$port_number/tcp --permanent
  - sudo firewall-cmd --reload
- Ubuntu
  - sudo ufw allow $port_number
On /camera, the first user to register a camera will need to enter an IP address for the camera to connect to. We provide some suggestions, but you can also input any valid IP address. We recommend checking tools like ifconfig. If the entered IP doesn’t work, you can click “Edit WiFi Settings” to modify it. You'll also need to provide the WiFi settings for the network that the on-prem machine is connected to.
Subsequent registrations will use the IP address from the initial registration and will not allow you to change it. However, a super user can change the IP address at /matroid-live/ip_address, or by going to the Control Panel, then IP address.
If your machine's IP address changes for any reason, a super user must change the entered IP address in the control panel. Any registered cameras will need to be re-registered to their streams with new QR codes. For updated IP address suggestions in the camera setup form, run the following command:
- sudo systemctl restart matroid

How does the provided monitoring work?

The default monitoring for a new camera stream detects people and cars using Matroid's Everyday Objects detector. You can change which objects are being detected in the monitoring settings, or make a new one with any detector that you'd like.

Because monitorings require credits, we provide 100,000 credits for the first successful registration of each kit. These credits - enough to monitor a stream for 3 days - are bound to the account on which you first registered the camera. To view your credit usage, click on your name in the upper right corner, select Account, then Credits.

How do I manually reset the camera or update its firmware?

You can reset the camera through the associated stream's settings. There are two ways to open stream settings from the Streams Dashboard (/streams):

Click on the stream of interest to reach the stream overview, then click Settings in the sidebar.
Click the hamburger dropdown icon on the stream thumbnail, then click Settings.

At the bottom of the settings, you should find a “Reset Camera” button. Once this is clicked, the camera information will be erased, except for the WiFi details which are used to generate the QR code (which is also accessible through the hamburger dropdown). The camera will then automatically reset after a few minutes, or you can physically reboot the camera by unplugging the power cable and plugging it back in. After the camera powers back on, it will begin the reset process. The camera can now be registered to a different stream - either a new one at /camera or an existing one by opening the QR code from the stream's dropdown menu.

Note: if a camera cannot find its associated stream or connect to its stored WiFi network, it will automatically reset and reboot.

When new camera firmware is available, a badge will appear next to the Settings tab of the stream overview page. Click the “Update” button next to Firmware Version in the stream settings to begin the update process. Similarly to resetting, you can manually reboot the camera to trigger the update.

Similarity search

What is similarity search?

Matroid Similarity Search lets you build visual search indexes over vast amounts of image and video data and search those indexes for content similar to a query image. You can search your collection using these indexes in milliseconds. It’s similar to reverse image search with 2 key differences:

The collection is your data
You can search with custom indexes that focus on the image characteristics you care about, like facial or specific object features

After creating a collection, you can choose a detector to build a search index with. When the search index is ready, you can search your collection for media similar to another query image. You can also search for media in your collection with detector scores above specified thresholds.

How do I create a similarity search collection?

What sources can a collection be created from?

How do I create a new search index?

Which detector should I create a search index with?

In general, you should choose a detector with labels similar to the things you’d like to search your collection for. If you’re going to search your collection for specific faces, you should index it with a face detector. If you’re going to search for specific cars, index with a detector that can distinguish between different makes and models of cars.

If you want to find all the instances of a particular object in your collection, you should index with a detector that has a label for that object. For example, if you were looking for airplanes, you could index your collection with a detector that has an airplane label. Then you could search your collection for all media that had an airplane score above a particular threshold.

How do I search my collection using a query image?

To search your collection for images similar to a query image, first go to matroid.com/similarity-search, click on the collection you want to search, and select the index you’d like to search again. You should choose an index tuned for features that you'd like to search for. E.g. If you want to find faces in your collection similar to a face in another image, choose an index created with a face detector. For a more generic object, choose an index created with a classifier.

Next select the tab relevant to the type of image you’d like to query with: choose “By image” for a local image file, “By webcam snap” for a webcam snapshot, or “By video clip” to select a frame from a local video.

Once you’ve entered a query image, you can select the specific part of the image you’d like to find similar matches for. If your index was created with a natural imagery detector, you can drag a box over the part of the image you’d like to query with:

If your index was created with a face detector, you can select the pre-localized face you’d like to find similar matches for:

After selecting a part of an image to query with, the most similar media in your collection should appear shortly.

How do I search my collection by detector score?

To search your collection for images by detector score, first go to matroid.com/collections, click on the collection you want to search, and select the index you’d like to search again. Next, select the “By detector scores” tab. Finally, select the labels and thresholds you’d like to search with:

Note that you will be shown results that have a score above any of the specified thresholds. I.e. In the example above, you’d be shown any faces with a score of 83% or higher for Channing Tatum or a score of 90% or higher for Jennifer Lawrence.

Which files in a collection get indexed?

All video and image files in your collection will be indexed. Note that images and videos from compressed files (like .zip or .tar files) will not be extracted.

A note on S3 prefixes: all images and videos with the S3 prefix you specify when creating your collection will be indexed. Say you have a bucket with the following structure:

s3://my-bucket/videos/
s3://my-bucket/videos-test/

If you create a collection with the S3 URL s3://my-bucket/videos/, only the s3://my-bucket/videos/ will be indexed. If you create a collection with the S3 URL s3://my-bucket/videos, both s3://my-bucket/videos/ and s3://my-bucket/videos-test/ will be indexed, because they both start with s3://my-bucket/videos.

What happens if I delete files from my collection?

Request A Demo Today

Let us show you how the power of computer vision can transform your organization

Request A Demo