Few Shot Object Detection without Fine-Tuning

Last decade has witnessed a series of breakthroughs in artificial intelligence field which led to the emergence of high-quality object detection algorithms. However, to learn a new concept, conventional methods require a complex process involving collecting and labelling data, model modification, developing fine-tuning strategy. In this work, we address the task of detecting objects on novel categories without fine-tuning given a few labelled data. We propose a simple yet effective model architecture consisting of fully-convolutional neural network which extracts features and transformer model which matches extracted features and features from few labelled examples. Proposed model effectively utilizes information about few labelled examples and performs single-stage detection. To train and evaluate our model we use custom synthetic dataset.