GoodVision uses a default classification scheme consisting of 8 basic classes - cars, vans, buses, trucks, heavy trucks, motorcycles, bicycles, and persons (see the full description here). However, the users can set up a customized class scheme according to their needs and standards. The object classes can be renamed, merged together, or completely removed from the scheme. You can also add new classes by applying coefficients to the remaining classes (explanation below). Remember, the adjusted classification will then be used through the whole Project and cannot be edited in a later stage anymore.
Picture 1: Possibility of adjusting the classification scheme when creating a new Project
A click on the “Create custom scheme” button will open a new window where the default scheme can be customized and the standard object classes manipulated. Let’s explore what the possibilities are.
Picture 2: Possibilities when adjusting the classification scheme
Each custom scheme has to have a name. The adjusted scheme will be saved on your profile under the scheme name, and you will then be able to use it for other potential projects later.
All default classes can be renamed (e. g. to your language), even those newly added by you.
By clicking on the percentage icon behind the traffic object names, you can adjust the coefficients (weights) for each class to adjust their distribution in the dataset, merge them together or delete them completely by clicking on the cross icon. For example, we can create a simple 4-objects scheme with no persons and bicycles (when we are interested in vehicle traffic only for the given Project), and where all vans are merged to cars and all heavy trucks to trucks.
Picture 3: Editing class coefficients
Similarly, we could move just a part of one class to another one. That might be useful in those regions where a part of the cars is considered rather as vans (like pick-ups, 7-seaters, etc.) or when transport vans are classified as trucks rather than seater vans.
Picture 4: Editing class coefficients
The most advanced step is to create new classes from the default scheme. You can create a new object class by clicking on the “Add new class” button. Name your new class, choose the current class/es from which you want to build the new class, and assign the percentage representation of the original classes. The new class will be extracted from the original set of object classes.
Picture 5: Adding new classes to the classification scheme
In this picture 5 example, we have created a scheme where 5% of cars and 2.5% of vans are to be considered as "taxis" (a newly added class), 50% of buses as "electro buses" and a new class "scooter" is created from 20% of motorbikes. These coefficients will be then used for every video footage uploaded and processed in the given Project. They will be visible both in the data in widgets and reports and in the form of the anonymized colored trajectories on the describe page of each Data Source.
Important note: Please keep in mind that adjusting the scheme this way, i. e. using the coefficients to split or add new traffic object classes, happens just on the statistical level. That means that the traffic volumes will be based on the chosen weights but will still originate from the initial 8-class dataset. Normally, clients use some additional research for setting up the coefficients in a realistic way.
GoodVision offers that service as well, so feel free to contact us in case you need to quickly set the class coefficients based on some manual verification from recorded footage.
Another, more robust way is to train the GoodVision neural network to recognize completely new or different object classes directly when processing the video. This option is qualitatively high but requires some additional costs and time for preparation. Moreover, it needs training and probably also some data input - recordings with some thousands of pictures (or dozens of hours) of the desired object class in different conditions/locations. Anyway, some GoodVision clients already use this possibility of an enriched classification scheme based on network training to deliver more precise results for the regional surveys. Please let us know if you are interested in this feature and we will prepare a description tailored to your needs.