Some of the technology we use is not compatible with every device. Rather than frustrate the user with bugs, we've disabled features for certain devices until those features can be supported. You can get the full experience on your desktop browser. Chrome on Android is also well supported.
The power of matroid is that you can make your own detector to detect anything you desire. From the search page, you just enter the labels you wish to detect in the search box.
If you are already working on a detector, labels you’ve previously provided will show above the search box and entering labels will add to your detector. Or you can return to the detector training page by clicking “Pending detector” in the header navbar. (Note: you can’t enter the same label twice.)
Matroid is striving to optimize the training for particular detectors that may demand extra detail detection. If you are interested in detecting certain people’s faces, create labels associated with each face you want to detect, add images that clearly show the face (using the bounding box tool if necessary - see the next question).
Click the tab on the training page. You can upload image files directly from your computer, select frames from a local video file, or select screenshots from your ip camera (see registering your ip camera).
The scores for each label at the timestamp represent the highest score that particular label had at the given timestamp. If a given timestamp has a high score for the label "Person wearing glasses", that means the timestamp most likely has someone wearing glasses in it. It's possible for the same timestamp to have a high score for "Person without glasses", which would mean there is a different face in the picture not wearing glasses.
In the video player, only the absolute highest score is shown, but the JSON or CSV results have all the scores for each timestamp included.
A good place to start is by looking at the images you've gathered, since the most common issues we see with training come from the data that's being fed in.
1. Some tips for creating a good detector:Choose label images must have enough diversity
The detector creation process is designed to train the detector to pick up on the common attributes among different images in each label. Making sure that the images supplied have only the distinguishing attributes as common and carefully weed out the wrong image labels before detector creation. For example if you have all images of cups with a red background, and all images of spoons with a white one, then the detector will end up learning to predict labels based on the background color. To avoid this, try to take diverse examples with
2. Avoid building broad category labels
Consider splitting big categories that cover a lot of different physical forms into smaller labels that are more visually distinct. For example instead of 'vehicle' you might use 'car', 'motorbike', and 'truck'.
3. Try training the detector with large number of examples per label that you are likely to encounter in your application
For the detector to perform well, we recommend having at least a hundred images/snapshots for each kind of label you want to detect. The more examples for each label that you add, the better the performance of your detector is likely to be. Also, make sure the examples are a good representation of what your application will actually encounter. For example, if you want to detect faces of people at your doorstep then you should make sure that labels have some such examples under different lighting conditions.
Matroid detectors in general have the capability to distinguish a large variety of subjects but you might occasionally encounter cases where your detector doesn’t perform well for your particular labels. Please contact us with your specific issues and we will work to make your detectors more powerful.
You may make a request to publish any of the Matroid detectors you have created. If the publication request is approved, your detector will appear in Matroid’s public detectors.
To publish one of your detectors, go to the My Detectors page and click on the detector’s name. On the next page, click the blue “Publish Detector” button and then fill out a name, description, and categories for your detector. This information will help other Matroid users find your detector.
Note that 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%. Learn more.
Once you make a publication request, 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.
Visit the pricing page to sign up for one of our subscription plans. Alternatively, if you publish high quality Matroid Detectors that make it to the public gallery, you can receive additional credits.
Positive images (examples of the label, e.g., a lion picture for the label "lion")
Negative images (examples of things that are not the label but could be confused with the label e.g., a tiger picture for the label "lion")
Note that when you redo an existing detector with improved data from feedback and other data sources, you are charged for all the images used, not just the new ones.
During training, if you realize you've made a mistake, you can delete or redo the detector from the training page. If you do, the training process will end and Matroid will refund the credits you spent to train.