Designed and developed a custom end-to-end model training platform for real-time instance segmentation and object detection tasks that featured:
- better organized raw data sourcing and storage
- improved standards for data preparation and labeling
- labeled data organization
- custom data augmentations and transformations
- platform architecture based on modularity, abstraction, encapsulation and reusability
- improvements to the instance segmentation and object detection model architecture
- configuration file organization with all the important parameters in a single place
- custom metrics and automated evaluation pipeline
- clean and straightforward pipeline architecture, removing any unnecessary and harmful steps
Featured are predictions on synthetic data generated by DALL-E. The model was not trained on such images, so the examples illustrate model’s generalizability to images, styles it has not seen before.
As part of this work, the following improvements to the instance segmentation model were achieved:
- major improvements to the model performance and generalizability across number of sites, products and customers
- optimized inference time (x2) and gpu memory consumption during inference (x3)
- reduction in dockerized model image size (x3)
Akbota Anuarbek 