Computer Vision Software Development Services

It will depend on the complexity of the project and the selected model type. The approximate numbers are:

1. Around 500 – 2000 annotated images per class: for basic tasks such as simple object or barcode recognition. 
2. Around 5,000 – 10,000 annotated images: for intermediate use cases, like recognition of multiple objects or working with cluttered backgrounds.
3. From 20,000 annotated images: for complex tasks, especially in such domains as healthcare and fintech (medical imaging, smart surveillance).

But no worries – if you don’t fit in the range, we can figure out a solution. For example, we can start with pre-trained models or few-shot learning and then gradually expand.

On average, it might take from 3 to 6 months to develop a full-featured computer vision solution. Here is what the process looks like when you partner with SoftTeco:

1. Problem definition & planning (1-3 weeks): the first thing that we’ll do is the definition of KPIs and success criteria and the assessment of your current infrastructure as well as model selection.
2. Data collection & annotation (2-6 weeks): we will collect high-quality data, annotate the generated dataset, and set up a data labelling tool.
3. Model training & setup (3-5 weeks): our experts train the model on the collected annotated data and validate its performance.
4. Evaluation & testing (3-4 weeks): we evaluate the accuracy of delivered results, test the model on unseen data, and optimize it for false positives or negatives.
5. Deployment (2-4 weeks): once we reach the expected level of model’s performance, we integrate it with your current system, configure APIs, and deploy the model into the production environment.
6. Optimization & monitoring (ongoing upon request): once the model is deployed, our work doesn’t stop. We offer post-launch support and maintenance, monitor the model’s performance, provide needed documentation, and retrain the model on new data.

SoftTeco has an established in-house framework for testing and validating the accuracy of our computer vision models. It includes several processes. We use a dedicated validation dataset, isolated from the testing and training datasets, and ensure that it contains real-life edge use cases. We also define evaluation metrics for each task. The KPIs for segmentation, for example, might include such metrics as IoU, pixel accuracy, dice coefficient.

To ensure the model performs equally well in the testing and production environment, we always perform field validation aka real-world testing. We also use A/B testing for model optimization and use a logging system to capture those cases where the model underperforms. This multimodal approach to model testing helps us ensure the accuracy and relevance of results.

Yes, we can – but note that it will require a few extra steps to take before we collect your training data.

We can start by planning and implementing the data collection process. This includes definition of available data sources and setup of data pipelines. Once the first batch of data is collected, we can start data annotation and MVP model training.

While working on the data pipelines setup, we can use public datasets or synthetic data to train a baseline computer vision model. This will help us quickly validate your idea and test its feasibility before working with real data and a larger model.