A study on transformer-based Object Detection

Abstract

This paper focuses on transformers based end-to-end object detection methods. End to end object detection is a new paradigm that has got attention in recent times. It does not require complex hand-engineered components such as non-max suppression to detect objects inside an image. Various methods are proposed to date to enhance fully end-to-end object detectors, most of them are based on the attention mechanism. In this work, we analyze some algorithms which involve transformers for the purpose of object detection. We discuss end-to-end models in which we have focused on Adaptive clustering-based transformers which solve attention encoder redundancy, Deformable Detection Transformers in which the attention module attends a limited collection of key sampling points, Unsupervisedly pre-trained Detection Transformers which are pre-trained on random query patches from the given image to improve accuracy and finally the Transformer-based Set Prediction using FCOS. These enhanced models not only improve the mean average precision of the model but also improves the total convergence time.