Human-Centric Perception for Child Sexual Abuse Imagery

Camila Laranjeira; João Macedo; Sandra Avila; Fabrício Benevenuto; Jefersson A. dos Santos

cs.CVMar 28, 2026

Human-Centric Perception for Child Sexual Abuse Imagery

Authors: Camila Laranjeira, João Macedo, Sandra Avila, Fabrício Benevenuto, Jefersson A. dos Santos

Organizations: Universidade Federal de Minas Gerais (UFMG), Brazil. · the Federal Police of Brazil. · Universidade Estadual de Campinas (UNICAMP), Brazil. · the University of Sheffield, United Kingdom.

Abstract

Law enforcement agencies and non-gonvernmental organizations handling reports of Child Sexual Abuse Imagery (CSAI) are overwhelmed by large volumes of data, requiring the aid of automation tools. However, defining sexual abuse in images of children is inherently challenging, encompassing sexually explicit activities and hints of sexuality conveyed by the individual's pose, or their attire. CSAI classification methods often rely on black-box approaches, targeting broad and abstract concepts such as pornography. Thus, our work is an in-depth exploration of tasks from the literature on Human-Centric Perception, across the domains of safe images, adult pornography, and CSAI, focusing on targets that enable more objective and explainable pipelines for CSAI classification in the future. We introduce the Body-Keypoint-Part Dataset (BKPD), gathering images of people from varying age groups and sexual explicitness to approximate the domain of CSAI, along with manually curated hierarchically structured labels for skeletal keypoints and bounding boxes for person and body parts, including head, chest, hip, and hands. We propose two methods, namely BKP-Association and YOLO-BKP, for simultaneous pose estimation and detection, with targets associated per individual for a comprehensive decomposed representation of each person. Our methods are benchmarked on COCO-Keypoints and COCO-HumanParts, as well as our human-centric dataset, achieving competitive results with models that jointly perform all tasks. Cross-domain ablation studies on BKPD and a case study on RCPD highlight the challenges posed by sexually explicit domains. Our study addresses previously unexplored targets in the CSAI domain, paving the way for novel research opportunities.

Figures & tables

Reference	Year	Main Task	Auxiliary Task
[ 21 ]	2010	Nudity Classification	Skin Segmentation
[ 24 ]	2011	CSAI Classification
[ 25 ]	2014	CSAI Classification	Skin Segmentation
[ 18 ]	2014	Nudity Classification	Skin Segmentation
[ 18 ]	2014	Child Classification	Face Detection
[ 22 ]	2016	CSAI Classification	Skin Segmentation

TABLE I : Overview of CSAI classification methods. Main tasks are features or models proposed by the authors, while auxiliary tasks involve off-the-shelf pre-processing methods.

Reference	Year	Targets (Task)
[ 26 ]	2017	Person (C), Pose (C), Action (C), Age (C), Scene (C), Attire (C), Face expression (C), Illumination (C).
RCPD [ 5 ]	2018	Person (D), Face (D), Nude private parts (D), Age (C), Gender (C).
Majura [ 6 ]	2018	Pornography (C), Nudity (C), Participants (C), Penetration (C), Props (C), Virtual (C), BDSM (C), Bodily Fluids (C).

TABLE II : Overview of CSAI labeling schemes in the literature. C: Classification, D: Detection.

Fig. 1 : Examples of labeled samples from BKPD. Top row: sourced from SOD, bottom row: sourced from Open Images. Boxes are color-coded per associated individual.

Dataset	person	head	chest	hip	l-hand	r-hand
SOD	1121	951	833	1013	767	815
Open-Images	2303	2180	1693	1238	1199	1274
Total	3424	3131	2526	2251	1966	2089

TABLE III : Total number of annotated parts on BKPD.

Fig. 3 : Samples with annotations from COCO-Keypoints (left) and COCO-HumanParts (right).

Class	CSAI	CSAI	CSAI	CSAI	Porn	Porn	Other	Other	Other	Other
Age	$>$ 13	$>$ 13	$\leq$ 13	$\leq$ 13	$>$ 13	$>$ 13	$>$ 13	$>$ 13	$\leq$ 13	$\leq$ 13
Gender	M	F	M	F	M	F	M	F	M	F
Person	79	8	382	581	82	294	76	74	226	210
Face	6	6	324	451	75	280	76	74	226	209
Breast	6	4	120	193	26	290	17	30	52	17
Genital	75	5	264	330	49	165	0	0	0	0

TABLE IV : Number of detection labels in RCPD.

Method	Backbone	Size	AP	AP 50	AR	AR 50
Associate Embedding [ 45 ]	ResNet50	512	46.6	74.2	56.6	81.0
Associate Embedding [ 45 ]	ResNet101	512	55.4	80.7	62.2	84.1
Associate Embedding [ 45 ]	ResNet152	512	59.5	82.9	65.1	85.6
Associate Embedding [ 45 ]	Hourglass	512	61.3	83.3	65.9	85.0
HigherHRNet [ 46 ]	HRNet-W32	512	67.7	87.0	72.9	89.0
HigherHRNet [ 46 ]	HRNet-W48	512	68.6	87.3	73.1	89.2

TABLE V : Pose estimation results for bottom-up methods on COCO2017 val set. SOTA results are presented as reported for single-scale testing. Our methods are trained on the joint COCO-Keypoints-HumanParts dataset. *BKP-Association’s contribution is the association algorithm, which does not impact the pose estimation task.

Fig. 4 : Qualitative results for pose estimation for small and tiny persons. Original images and zoom-ins side-by-side.

Method	All categories							Per category AP
Method	AP	AP 50	AP 75	AP T	AP S	AP M	AP L	person	head	face	r-hand	l-hand	r-foot	l-foot
Faster-C4-R50 [ 49 ]	32.0	55.5	32.3	9.9	38.5	54.9	52.4	50.5	47.5	35.5	27.2	24.9	19.2	19.3
RetinaNet-R50 [ 50 ]	32.2	54.7	33.3	10.0	39.8	54.5	53.8	49.7	47.1	33.7	28.7	26.7	19.7	20.2
Faster-FPN-R50 [ 51 ]	34.8	60.0	35.4	14.0	41.8	55.4	52.2	51.4	48.7	36.7	31.7	29.7	22.4	22.9
Hier-R50 [ 43 ]	36.8	65.7	36.2	19.9	41.9	53.9	47.5	53.2	50.9	41.5	31.3	29.3	25.5	26.1
Faster-FPN-R101 [ 51 ]	36.0	62.1	36.5	14.8	43.0	57.2	54.8	52.6	49.3	36.9	33.3	30.3	24.3	24.4

TABLE VI : Detection results on COCO-HumanParts val set. SOTA results are presented as reported for single-scale testing. Our methods are trained on the joint COCO-Keypoints-HumanParts dataset.

Method	Part	AP 50	JAP	a CA	Part	AP 50	JAP	CA
Hier-R50 [ 43 ]	H	78.0	73.4	93.1	F	68.7	64.8	93.5
BPJDet-L [ 15 ]	H	82.5	76.8	95.6	F	71.2	64.9	94.2
BKP-Assoc.	H	81.6	74.8	94.5	F	72.1	65.1	93.7
YOLO-BKP	H	82.9	79.5	95.7	F	76.9	73.8	95.8
Hier-R50 [ 43 ]	LH	49.6	47.6	94.0	RH	52.3	50.1	93.8
BPJDet-L [ 15 ]	LH	65.5	61.4	96.0	RH	65.0	60.6	95.8

TABLE VII : Association results on COCO-HumanParts val set. H: Head, F: Face, LH: Left Hand, RH: Right Hand, LF: Left Foot, RF: Right Foot.

Fig. 5 : Qualitative results on COCO2017 val set. Left: BKP-Association, right: YOLO-BKP. Bounding box colors indicate association to the same individual.

Method	BKPD-OI				BKPD-SOD
Method	AP	AP 50	AR	AR 50	AP	AP 50	AR	AR 50
DEKR-w32 [ 47 ]	85.3	93.7	90.7	97.9	67.8	83.1	72.7	87.7
DEKR-w48 [ 47 ]	83.5	91.8	91.4	98.3	67.7	82.9	73.7	89.4
YOLOv7-w6-pose [ 13 ]	91.6	96.6	93.9	98.0	68.9	85.1	75.0	87.7
BKP-Assoc.	88.9	95.6	91.9	97.6	67.6	84.3	72.9	86.6
YOLO-BKP	88.5	96.0	91.9	98.0	66.0	83.9	73.5	87.6

TABLE VIII : Pose estimation results on BKPD for bottom-up methods trained on COCO-Keypoints, except our methods, which are trained on the joint COCO-Keypoints-HumanParts.

Method	Size	All categories						Per category AP
Method	Size	a AP	AP 50	AP 75	AP S	AP M	AP L	Person	Head	a LH	RH
	BKPD-OI
Hier-R50 [ 43 ]	1536	50.2	79.0	52.5	23.1	40.6	59.0	70.8	65.3	31.6	33.0
BPJDet-L [ 15 ]	960	61.9	86.6	66.5	23.4	48.7	70.8	83.0	72.9	44.5	47.3
BKP-Assoc.	960	62.7	88.2	67.2	24.9	48.5	71.8	84.3	73.4	45.3	47.8
YOLO-BKP	960	59.1	87.4	61.3	21.0	47.2	65.9	85.8	71.4	39.6	39.7

TABLE IX : Detection results on BKPD for models trained on COCO-HumanParts, except our methods, which are trained on the joint COCO-Keypoints-HumanParts.

Method	Size a	Head			L-Hand			R-Hand
Method	Size a	AP 50	JAP	a CA	AP 50	JAP	a a CA	AP 50	JAP	CA
	BKPD-OI
Hier-R50 [ 43 ]	1536	93.5	88.8	95.5	66.5	62.5	93.7	68.1	63.6	92.8
BPJDet-L [ 15 ]	960	95.5	91.5	97.4	78.1	73.4	96.0	79.2	73.0	94.8
BKP-Assoc.	960	97.1	92.0	96.9	77.2	68.6	92.4	78.5	70.6	92.7
YOLO-BKP	960	97.6	95.1	97.4	78.4	76.1	96.6	77.6	75.0	96.8

TABLE X : Association results on BKPD for models trained on COCO-HumanParts, with the exception of our methods, which are trained on the joint COCO-Keypoints-HumanParts.

Method	Tune	Val	Head				Chest				Hip				L-Hand				R-Hand
Method	Tune	Val	AR 50	AP 50	JAP	a CA	AR 50	AP 50	JAP	a CA	AR 50	AP 50	JAP	a CA	AR 50	AP 50	JAP	a CA	AR 50	AP 50	JAP	CA
BKP-Assoc.	All	OI	95.2	96.2	91.0	96.7	62.6	65.1	62.1	97.5	65.5	64.8	62.5	97.9	78.3	78.1	68.8	91.8	77.3	74.9	64.0	92.2
YOLO-BKP	All	OI	97. 5	97.1	95.4	97.9	83.2	76.7	75.1	98.1	84.3	76.1	75.7	99.6	77.0	70.0	68.3	97.8	63.9	65.3	64.3	97.0
BKP-Assoc.	SOD	OI	88.5	93.6	84.9	94.1	34.9	23.6	21.6	94.6	44.6	39.5	38.5	98.2	72.0	71.7	62.1	91.8	71.6	70.2	58.4	90.8
YOLO-BKP	SOD	OI	94.7	93.8	92.1	97.6	68.0	54.5	53.4	97.9	69.7	57.3	57.1	99.3	71.1	63.4	62.1	96.7	65.8	58.0	56.7	97.1
BKP-Assoc.	All	SOD	94.4	95.1	86.7	95.2	59.8	60.7	58.0	96.7	69.9	72.4	69.5	96.7	76.9	76.9	70.7	93.9	82.3	80.7	75.5	94.8

TABLE XI : Detection and association performance with various data splits from BKPD.

Fig. 6 : Example hierarchic detections from BKP-Association (left) and YOLO-BKP (right).

Fig. 7 : Comparison of inferences and ground-truth on RCPD. Red : detection, green : ground-truth. Gray traces refer to pose estimation results. Left: person, Right:face/head.

Method	Domain	Tune	#People $=1$				#People $\geq 2$
Method	Domain	Tune	AP	AP 50	AR	AR 50	AP	AP 50	AR	AR 50
BKP-Assoc.	CSAI	–	30.3	66.1	44.3	76.7	27.3	60.6	38.7	69.1
YOLO-BKP	CSAI	–	29.6	65.1	46.0	79.5	28.8	63.3	41.0	73.8
BKP-Assoc.	CSAI	BKPD	30.2	67.6	44.1	77.8	27.7	62.9	38.4	70.4
YOLO-BKP	CSAI	BKPD	29.8	66.3	47.6	81.4	27.6	63.3	41.9	77.3
BKP-Assoc.	Porn	–	34.3	78.7	50.5	89.4	23.3	58.5	37.3	72.3

TABLE XII : Detection summary of results with the standard IoU metric for the class person on RCPD.

Method	Domain	#People $=1$				#People $\geq 2$
Method	Domain	AP ${}_{60}^{in}$	AP ${}_{75}^{in}$	AR ${}_{60}^{in}$	AR ${}_{75}^{in}$	AP ${}_{60}^{in}$	AP ${}_{75}^{in}$	AR ${}_{60}^{in}$	AR ${}_{75}^{in}$
BKP-Assoc.	CSAI	96.7	94.4	99.0	97.5	81.2	80.5	83.7	82.9
YOLO-BKP	CSAI	94.1	91.6	99.2	98.1	81.0	80.0	83.4	82.5
BKP-Assoc.	Porn	99.2	98.6	100	99.5	84.8	83.9	88.1	87.6
YOLO-BKP	Porn	97.9	95.3	99.5	97.5	86.2	85.8	89.2	89.2
BKP-Assoc.	Other	98.6	98.0	100	99.6	89.6	89.5	94.1	94.1

TABLE XIII : Person detection results for models fine-tuned on BKPD, evaluated with inner IoU.

Fig. 8 : Example detections of private parts on RCPD. Detections in red and ground-truths in green .

Method	Domain	#People $\leq 1$				#People $\geq 2$
Method	Domain	AR ${}_{50}^{in}$	AP ${}_{50}^{in}$	JAP ${}_{50}^{in}$	CA	AR ${}_{50}^{in}$	AP ${}_{50}^{in}$	JAP ${}_{50}^{in}$	CA
		Head \| Face
BKP-Association	CSAI	99.0	97.9	97.9	100	95.1	92.2	90.2	98.9
YOLO-BKP	CSAI	98.4	95.5	95.2	98.8	95.8	92.0	90.6	97.8
BKP-Association	Porn	98.9	97.6	97.2	99.5	96.2	92.2	91.4	99.3
YOLO-BKP	Porn	99.5	98.2	97.7	98.4	96.8	93.9	92.9	98.0

TABLE XIV : Detection and association performance of RCPD human parts, evaluated with inner IoU.

Keypoint Detection

Keypoint detection locates distinctive image points or meaningful landmarks for matching, pose estimation, tracking, and geometric localization.

See all 28 papers

CSA-Graphs: A Privacy-Preserving Structural Dataset for Child Sexual Abuse Research
Apr 8, 2026Carlos Caetano, Camila Laranjeira, Clara Ernesto +5Scene Graphs Image Classification
Universidade Estadual de Campinas (UNICAMP), Brazil · Instituto Federal de Educação, Ciência e Tecnologia de Minas Gerais (IFMG), Brazil · Universidade de São Paulo (USP), Brazil +3
Classifying by Proxy: Explainable and Reproducible Ensemble of Proxy Tasks for Child Sexual Abuse Imagery Classification
Jun 14, 2026Clara Ernesto, Carlos Caetano, Sandra Avila +3Explainable Artificial Intelligence Ensemble Learning
Instituto de Ciências Matemáticas e de Computação (ICMC), Universidade Estadual de São Paulo (USP), Brazil · Instituto de Computação (IC), Universidade Estadual de Campinas (UNICAMP), Brazil · Departamento de Ciência da Computação, Universidade Federal de Minas Gerais (UFMG), Brazil +1
Synthetic-Child: An AIGC-Based Synthetic Data Pipeline for Privacy-Preserving Child Posture Estimation
Mar 3, 2026Taowen ZengHuman Pose Estimation Synthetic Data Generation
Independent Researcher

Apr 8, 2026J/K move · Enter open · S save

CSA-Graphs: A Privacy-Preserving Structural Dataset for Child Sexual Abuse Research

Carlos Caetano, Camila Laranjeira, Clara Ernesto, Artur Barros+4

Universidade Estadual de Campinas (UNICAMP), Brazil · Instituto Federal de Educação, Ciência e Tecnologia de Minas Gerais (IFMG), Brazil · Universidade de São Paulo (USP), Brazil+3

Scene Graphs Image Classification Privacy-Preserving ML

Open paper AskPDF arXiv

Child Sexual Abuse Imagery (CSAI) classification is an important yet challenging problem for computer vision research due to the strict legal and ethical restrictions that prevent the public sharing of CSAI datasets. This limitation hinders reproducibility and slows progress in developing automated methods. In this work, we introduce CSA-Graphs, a privacy-preserving structural dataset. Instead of releasing the original images, we provide structural representations that remove explicit visual content while preserving contextual information. CSA-Graphs includes two complementary graph-based modalities: scene graphs describing object relationships and skeleton graphs encoding human pose. Experiments show that both representations retain useful information for classifying CSAI, and that combining them further improves performance. This dataset enables broader research on computer vision methods for child safety while respecting legal and ethical constraints.

Abstract

Keypoint Detection

Explore similar work

CSA-Graphs: A Privacy-Preserving Structural Dataset for Child Sexual Abuse Research

Classifying by Proxy: Explainable and Reproducible Ensemble of Proxy Tasks for Child Sexual Abuse Imagery Classification

Synthetic-Child: An AIGC-Based Synthetic Data Pipeline for Privacy-Preserving Child Posture Estimation