CVAT: your own training dataset
For a company’s own recognition model, it needs a set of images in which the same cases are described in the same way. The photo folder itself does not specify where the object is, what class it belongs to, and who confirmed the signature. A local server with CVAT can be thought of as the workspace of the team that prepares such markup and transmits the accepted version to the model developer. The purchase begins with the result of this command, and the equipment is selected according to the agreed process.
Request a quote
First, describe what the future system is taught.
For example, a company collects photos of housings and connecting elements to train a model to find certain visible features. One project is enough to mark the entire picture by category, another needs the boundaries of each object. These options create different amounts of manual work and different training data requirements. It is useful to select one practical task in advance and determine which file the developer should get after the first step.
For the first materials, prepare a short marker instruction: class name, positive examples, controversial cases, and a rule for a poorly visible object. If one employee notes the entire body, and the other only the damaged area, the total set is heterogeneous. Fixing the rules on a few dozen consistent examples is easier than redoing the entire archive.
Project, tasks and a single dictionary
CVAT allows you to create tasks with images or videos, set labels and attributes, and then upload annotations. In the implementation project, this is the basis for a coordinated work process: what materials are received, who is appointed to the party and who accepts the result. The class dictionary is saved with the version so that the change in definition is visible to all participants.
| Decision prior to markup | A practical example | What's saved |
|---|---|---|
| Type of result | Image category or object area | Rule of choice of geometry |
| Class boundary | Pollution and scratch are treated separately | Examples and controversial cases |
| Accessible visibility | Part of the facility is closed. | Agreed uncertainty mark |
| Party acceptance | The second employee is reviewing the result | Corrections and drafting as adopted |
Materials
- Originals and origins
- Image/video batches
- Manual or agreed advance markings
Team decisions
- Class version and instruction
- Verification of disputes
- Adoption of the party
Accepted annotations + originals + verified export format
Manual inspection has a separate result
It is more convenient for the inspector to receive a limited batch with an understandable task than the entire archive without status. It looks at the location of areas, missed objects, class and desired attributes. Comments are returned to work with the reason, and the adopted version receives its own designation. This procedure can be included in the organization of the project regardless of the choice of additional automatic modules.
A separate set of examples is useful to use to harmonize the work of employees. She answers the question of whether the team understands the instructions the same way. The test sample of the future model plays a different role: it evaluates the result of training. The composition of these sets and the rules for their use are set separately, preserving the origin of each image.
Where can the pre-markup model help?
If there is a suitable model for your classes, you can agree on its application to new batches. The employee receives the proposed areas and edits them before the result is accepted. In the CVAT documentation, the presence of automatic markup is related to the composition of the deployed instance; specific components, models and rights of use are included in the quotation separately.
Comparison of a person’s full work time: how much quotation he left, what had to be fixed and how many objects were missed. Quickly obtained areas can be uncomfortable for the editor. For selection, it is more useful to process the characteristic batch and see its correction than to focus only on the number of frames that the model gives out per second.
Transmit data so that it can be used
Agree with the developer the export format, attributes, file names and link to the originals. The CVAT documentation describes its own format with images and annotations, as well as other upload options. Their capabilities vary. At the first stage, it is worth opening an export in the intended recipient and checking several areas, classes and attributes with the accepted markup.
For frames from one video, the original recording and belonging to it are saved. When forming the training and verification parts of the project, it is useful to take these groups into account so that very similar neighboring personnel do not replace the assessment on another material. The accepted batch remains reproducible: the source, instruction, annotations and the composition of the export are stored together.
How to choose a server and implementation composition
The storage of originals, uploading images to employees, preparing videos and a possible pre-markup model create different loads. To calculate, file sizes, the number of markers, the growth of the archive and the composition of automatic operations are needed. The accelerator is selected for an actually agreed model; conventional manual markup does not by itself define the requirement for a large VRAM.
Own quotation -- AI Server Manager 256Price and availability upon request. In the application, specify the task, example of materials, classes, number of employees, desired export and city of delivery. Equipment, CVAT deployment, access setting, model connection and maintenance are agreed on composition. The designation 256 does not define the memory or performance of the selected machine.
Questions before ordering
Do I need to order automatic markup immediately?
You can start with a manual process, a consistent vocabulary, and a proven export. It is advisable to evaluate the connection model by how much time the employee spends correcting the quotation on your materials.
Which version of the CVAT will be installed?
Editorial, version, deployment method and licensed functions are specified in the quotation. CVAT Online or Enterprise features are not automatically included in the Community or in the cost of the server.
Does the markup mean the model has already been trained?
This is the preparation of data for agreed further use. Training, evaluation on a separate material and implementation of the model are independent stages of the project.
Basic quotation
AI Server Manager 256
Price on request. The actual configuration, availability, programs and works are agreed in the quotation. The number 256 does not determine the amount of memory.
Request a quotePrimary sources: CVAT Tasks: task creation, labels, exports and optional automatic annotation; CVAT for image / video: native annotation format; CVAT official repository. Versions, editions, licenses and implementation composition are agreed for the selected delivery.
