Detectron-2 Model Retraining

This project (github page) revolves around utilizing Detectron2, Facebook AI Research’s cutting-edge library for detection and segmentation algorithms. The primary objective is model retraining, focusing on enhancing segmentation capabilities for specific objects – in our project it is human – within the popular CoCo Dataset.

The process begins with Docker image setup, ensuring seamless deployment and consistency across environments. Leveraging pre-built Docker images streamlines the subsequent stages, ensuring dependencies are managed efficiently.

The CoCo Dataset serves as the foundation for model training, providing a diverse range of annotated images. Filtering techniques are employed to isolate relevant categories, such as “person” and “sports ball”, optimizing the dataset for targeted retraining. Balancing instance distribution within these categories mitigates biases, enhancing model performance and accuracy.

Following dataset preparation, the project delves into segmentation model retraining using Detectron2. By fine-tuning existing models with the curated dataset, the aim is to improve segmentation accuracy specifically for the identified categories. This step involves meticulous configuration and parameter tuning to achieve optimal results.

Upon successful retraining, the retooled model is applied to both image and video inference tasks. Through inference, the model’s efficacy in segmenting target objects is evaluated, showcasing its practical utility in real-world scenarios.

Overall, this project encapsulates the iterative process of model refinement, from dataset curation to retraining and inference, leveraging the capabilities of Detectron2 to push the boundaries of segmentation performance.