Body-Keypoint-Part Dataset (BKPD)

The Body-Keypoint-Part Dataset (BKPD) comprises human-centric images, and semiautomatic labels hierarchically organized for each individual, including skeleton keypoints and bounding boxes for the person's body and the following human parts: head, chest, hip, left hand, and right hand. We curated a total of 1,985 safe images from domains related to Child Sexual Abuse Imagery (CSAI), including 983 non-sensitive images of people sourced from the Open Images dataset[1], and 1,002 sexually explicit images of adults from the Sexual Organ Detection (SOD) dataset[2]. Regarding pose keypoints, we follow the COCO-Keypoints representation, since it leverages one of the largest and widely used benchmarks. It includes five keypoints for the head, six for the arms (shoulders, elbows, and wrists), and six for the legs (hips, knees, and ankles). Furthermore, the choice of body parts follows from our literature review of CSAI metadata produced by child protection institutions and Law Enforcement Agencies (LEA), specifying sensitive body parts commonly involved in sexually explicit interactions. Dataset Structure images.zip: .jpg files with raw images annotations_hier_coco_formatted.json: COCO-formatted JSON file, following COCO-HumanParts structure of hierarchically organized labels. The field "bbox" refers to the (x, y, w, h) person bounding box, "keypoints" are (xi, yi, vi) vectors with coordinates and visibility of each ith skeleton keypoint, while the field "hier" are (x0j, y0j, x1j, y1j, v1) vectors with coordinates and visibility of each jth body part bounding box, respectively the head, chest, hip, left hand, and right hand. labels_multiclass_yolopose_formatted.zip: A single file for each sample containing the corresponding labels formatted according with YOLO-Pose's original work. Each line in a file comprises a vector (ci, x0i, y0i, x1i, y1i) for each ith bounding box, providing its category and normalized coordinates, followed by 17 (x,y,v) points referring to the coordinates and visibility of each keypoint. Only person annotations (ci = 0) contain informative keypoint coordinates, with the remaining classes providing zero vectors. classes = {0: 'person', 1: 'head', 2: 'chest', 3: 'hip', 4: 'l-hand', 5: 'r-hand'} labels_hier_yolobkp_formatted.zip: A single file for each sample containing the corresponding labels formatted according with our proposed YOLO-BKP. Each line in a file comprises a vector (0, x0i, y0i, x1i, y1i, idj, ci) for each bounding box i and individual j, containing the id of the individual along with normalized coordinates and the category for each ith body part. Person annotations (ci = -1) are followed by 17 (x, y, v) points referring to the coordinates and visibility of each keypoint. part_classes = {0: 'head', 1: 'chest', 2: 'hip', 3: 'l-hand', 4: 'r-hand'} References [1] Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., ... & Ferrari, V. (2020). The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7), 1956-1981. [2] Tabone, A., Camilleri, K., Bonnici, A., Cristina, S., Farrugia, R., & Borg, M. (2021, August). Pornographic content classification using deep-learning. In Proceedings of the 21st ACM symposium on document engineering (pp. 1-10).