In this project, the challenge of detecting objects from new categories without fine-tuning and in conditions of minimal labeled data was addressed.
To tackle this, a simple yet effective model architecture was proposed. It comprises two main components — a fully convolutional neural network extracting feature descriptions and a transformer-based model associating information from several annotated examples with the extracted representations for making predictions.
The proposed model leverages information from multiple annotated examples and performs one-stage detection. The model was trained and evaluated both on a custom synthetic dataset and FSOD dataset.
This project was completed by Stanislav Mikhaylevskiy and Vladimir Chernyavskiy. If you have any questions or suggestions regarding this project, please feel free to contact us.