Image captioning is a fundamental task connecting computer vision and natural language processing. Recent researches usually concentrate on generic image captioning or video captioning among thousands of classes. However, they can not effectively deal with a specific class of objects, such as pedestrian. Pedestrian captioning is critical for analysis, identification and retrieval in massive collections of data.

Therefore, in this paper, we propose a novel approach for pedestrian captioning with joint visual context. Firstly, a deep convolutional neural network (CNN) is employed to obtain the global attributes of a pedestrian (e.g., gender, age, and actions), and a Faster R-CNN is utilized to detect the local parts of interest for identification of the local attributes of a pedestrian (e.g., cloth type, color type, and the belongings).

Then, we splice the global and local attributes into a fixed length vector and input it into a Long-Short Term Memory network (LSTM) to generate descriptions. Finally, a dataset of 5000 pedestrian images is collected to evaluate the performance of pedestrian captioning.

Experimental results show the superiority of the proposed approach.