Success Story - en

Optimizing the facial recognition dataset

Motivation for launching the project by the customer:
the need to speed up the training of neural network models without loss of quality, or training in the same time to a higher quality. With subsequent testing of the approaches on the customer’s target tasks.

Description of the initial situation:
  • existing pipelines for training customer models take a long time to train;
  • an increase in the quality of models is associated with an increase in the complexity of the models, as well as an increase in the work time and cost of training the model;
  • the target uses customers' personal data.

Project goals:
  • assessment of the effectiveness of existing methods when transferring them to the target task area;
  • anonymization of data using data set optimization methods.

MIL Team solution:
conducting research in the field of Data Optimization with a concentration on data manipulation, augmentation, data distillation (increasing the information content of a single picture). Providing research code and recommended pipeline to speed up the model training process.

To build the model, the following were used: open image datasets used in publications.
Project results: under NDA
Customer: under NDA
Technology stack: PyTorch, DVC, PyTorch Lightning
Computer Vision Research
Made on