ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision
This work addresses the efficiency and expressive power limitations of existing Vision-and-Language Pre-training (VLP) models for researchers and practitioners by proposing a significantly faster architecture.
This paper introduces ViLT, a Vision-and-Language Transformer model that simplifies visual input processing by removing convolutions and region supervision. ViLT achieves competitive or better performance on downstream tasks while being tens of times faster than previous VLP models.
Vision-and-Language Pre-training (VLP) has improved performance on various joint vision-and-language downstream tasks. Current approaches to VLP heavily rely on image feature extraction processes, most of which involve region supervision (e.g., object detection) and the convolutional architecture (e.g., ResNet). Although disregarded in the literature, we find it problematic in terms of both (1) efficiency/speed, that simply extracting input features requires much more computation than the multimodal interaction steps; and (2) expressive power, as it is upper bounded to the expressive power of the visual embedder and its predefined visual vocabulary. In this paper, we present a minimal VLP model, Vision-and-Language Transformer (ViLT), monolithic in the sense that the processing of visual inputs is drastically simplified to just the same convolution-free manner that we process textual inputs. We show that ViLT is up to tens of times faster than previous VLP models, yet with competitive or better downstream task performance. Our code and pre-trained weights are available at https://github.com/dandelin/vilt.