Applying Positional Encoding to Enhance Vision-Language Transformers

Abstract

Positional encoding is used in both natural language and computer vision transformers. It provides information on sequence order and relative position of input tokens (such as of words in a sentence) for higher performance. Unlike the pure language and vision transformers, vision-language transformers do not currently exploit positional encoding schemes to enrich input information. We show that capturing location information of visual features can help vision-language transformers improve their performance. We take Oscar, one of the state-of-the-art (SOTA) vision-language transformers as an example transformer for implanting positional encoding. We use image captioning as a downstream task to test performance. We added two types of positional encoding into Oscar: DETR as an absolute positional encoding approach and iRPE, for relative positional encoding. With the same training protocol and data, both positional encodings improved the image captioning performance of Oscar by between 6.8% to 24.1% across five image captioning evaluation criteria used.