Transformer-based Convolution-free Visual Place Recognition

Over the last decade, convolutional neural networks have been a core element in the recent remarkable advances in machine learning, computer vision, and robotics. Vision transformers have recently demonstrated great success in various computer vision tasks, motivating a tremendously increased interest in their deployment into many real-world vision applications. However, until now, the number of successful applications of transformers in robotics is somewhat limited. This work presents an approach to visual place recognition using a vision transformer (ViT). ViT trained from scratch, and two pretrained ViTs in base and large versions have been finetuned on a target dataset. The features extracted by transformers have then been used in place recognition using a k-NN. Finally, contrastive learning has been performed to embed the features and improve recognition performance. The algorithm has been evaluated in a dataset for indoor place recognition comprising images with 6-DOF viewpoint variations. Experimental results demonstrate that considerable gain in recognition accuracy can be obtained by finetuned transformers in comparison to results achieved by CNNs.