Multilingual speakers all over the globe often switch between languages known as code-mixing during communication or the employment of two or more languages in a single statement. Most of the works on code-mixing have been accomplished on text classification, question answering, dialogue understanding, etc. However, no specific research has been carried out on code-mixed image captioning in the past. We present an English-Hindi code-mixed image captioning dataset. We experiment with a novel architecture utilizing Faster-R CNN and Transformer as encoder and decoder with a geometric attention mechanism and other baselines. The code mixed dataset is created from a manually annotated parallel corpus of the MS COCO dataset in English and Hindi. Experimental findings show that the proposed technique surpasses the baselines in code-mixed situations.