Supplementary material: Spatio-temporal Person Retrieval via Natural Language Queries

In this section, we provide the further details of the dataset statistics. The description length. We first analyze the description length (i.e., the number of words in a description). Figure 1 shows the distribution of the number of words in a description. We can see that our dataset contains various lengths of descriptions. The average length of descriptions in our dataset is 13.1. We also show the comparison of the average description length of our dataset to those of other datasets in Table 1. ReferIt [7] and Google RefExp [12] are the datasets of referring expressions, each of which is true of only a single region in an image. The descriptions in VisualGenome [10], MSR-VTT [20] and MSCOCO [2] focus on regions in images, whole images and videos, respectively. Even though a description in our dataset focuses on a single person, the average description length of our dataset is larger than not only those of the datasets of which descriptions focus on regions in images, but also those of the datasets of which descriptions focus on the whole images or videos. This implies that the descriptions in our dataset tend to contain more detailed information than those in other datasets. The number of annotated people in a clip. Figure 2 shows the distribution of the number of people who are annotated with bounding boxes and descriptions in a single clip. While many clips contain only one annotated person, some clips contain multiple annotated people. The number of occurrences of each high-frequency word. Figure 4 shows the number of occurrences of the most frequently occurring words (Stop words are excluded). We can see that high-frequency words involve various types of words such as colors, actions, clothes and places. Figure 5 shows the comparison of frequencies of words in Figure 4 between our dataset and VisualGenome. While the frequencies of words describing colors (e.g. black, white, blue and red) and people (e.g. man, woman, girl and boy) in our dataset are close to those in VisualGenome, the 10 15 20 25 30 35