SenBen: Sensitive Scene Graphs for Explainable Content Moderation

Fatih Cagatay Akyon; Alptekin Temizel

cs.CVApr 9, 2026

SenBen: Sensitive Scene Graphs for Explainable Content Moderation

Authors: Fatih Cagatay Akyon, Alptekin Temizel

Organizations: Graduate School of Informatics, METU, Turkiye · Ultralytics, Inc., United States

Abstract

Content moderation systems classify images as safe or unsafe but lack spatial grounding and interpretability: they cannot explain what sensitive behavior was detected, who is involved, or where it occurs. We introduce the Sensitive Benchmark (SenBen), the first large-scale scene graph benchmark for sensitive content, comprising 13,999 frames from 157 movies annotated with Visual Genome-style scene graphs (25 object classes, 28 attributes including affective states such as pain, fear, aggression, and distress, 14 predicates) and 16 sensitivity tags across 5 categories. We distill a frontier VLM into a compact 241M student model using a multi-task recipe that addresses vocabulary imbalance in autoregressive scene graph generation through suffix-based object identity, Vocabulary-Aware Recall (VAR) Loss, and a decoupled Query2Label tag head with asymmetric loss, yielding a +6.4 percentage point improvement in SenBen Recall over standard cross-entropy training. On grounded scene graph metrics, our student model outperforms all evaluated VLMs except Gemini models and all commercial safety APIs, while achieving the highest object detection and captioning scores across all models, at $7.6\times$ faster inference and $16\times$ less GPU memory.

Figures & tables

Split	Frames	Movies	Sens.	Gen.	Obj/fr	Attr/fr	Pred/fr
Train	9,999	95	4,999	5,000	3.42	8.09	2.50
Val	2,000	31	1,000	1,000	3.39	7.95	2.50
Test	2,000	31	989	1,011	3.41	8.04	2.52
Total	13,999	157	6,988	7,011	3.41	8.06	2.51

Table 1: Dataset split statistics of SenBen. Annotation density is averaged per frame.

Category	Tag	Train	Val	Test	Total
Immodesty	immodesty	734	177	155	1,066
	nudity	370	62	88	520
	nudity art	143	25	31	199
	nudity implied	371	61	68	500
Sexual	sex. suggestive	579	104	111	794
	kissing	248	47	49	344

Table 2: Per-tag frame counts across splits. Frames may carry multiple tags (avg 1.3 per sensitive frame).

Figure 2: Multi-task training architecture. DaViT vision features feed both a decoupled Q2L tag head (detached from the decoder) trained with Asymmetric Loss, and a fully fine-tuned encoder-decoder for four scene graph tasks. Suffix-based object identity ( :N ) replaces bounding box tokens in attribute and predicate outputs. VAR loss (Eq. 3 ) penalizes low recall on sensitive vocabulary tokens across decoder tasks.

Aspect	Recall Loss	Skeleton Recall	VAR (Ours)
Domain	Segmentation	Segmentation	Autoregressive
Unit	Per-class pixels	Thin structures	Vocab tokens
Formulation	Multiplicative	Additive	Additive

Table 3: VAR Loss in context of recall-focused losses.

System	$\mathrm{R}_{\mathrm{SB}}$	$\text{F1}_{\mathrm{SB}}$	$\text{F1}^{\mathrm{tag}}$
CE (baseline)	.349	.389	.544
+Suffix	.366	.406	.509
+VAR	.376	.408	.509
+MinPermCE	.386	.415	.532
+Label smooth.	.385	.419	.533
+Sched. samp.	.392	.422	.516

Table 4: System ablation study.

\mathrm{R}_{\mathrm{SB}}

: SenBen Recall,

\text{F1}_{\mathrm{SB}}

: SenBen F1,

\text{F1}^{\mathrm{tag}}

: macro tag F1 (

\mathrm{R}_{\mathrm{SB}}

\text{F1}_{\mathrm{SB}}

category-macro-averaged).

Removed	immod	sexual	viol	subst	other	avg
$-$ Suffix	$-4.6$	$-4.5$	$-6.5$	$-2.9$	$-0.4$	$-3.8$
$-$ VAR	$-5.7$	$-7.7$	$-2.0$	$-0.9$	$-0.9$	$-3.4$
$-$ LS	$-3.3$	$-5.7$	$+0.3$	$+0.4$	$+0.8$	$-1.5$
$-$ SS	$-1.6$	$-2.8$	$+0.3$	$+0.8$	$-0.4$	$-0.7$
$-$ MinP	$+0.5$	$-2.4$	$-1.4$	$+1.7$	$-0.4$	$-0.4$

Table 5: Per-category

\mathrm{R}_{\mathrm{SB}}

change when removing each ingredient from the full decoder (VAR+SS+MinP+LS+Suffix). Values are

\Delta\mathrm{R}_{\mathrm{SB}}

in percentage points.

Model	Params	$\text{F1}^{\mathrm{tag}}$	$R^{\mathrm{obj}}$	$\mathrm{R}_{\mathrm{SB}}$	$\text{F1}_{\mathrm{SB}}$	cap
Gemini 3 Pro (low reas.)	—	.806	.295	.652	.647	.642
Gemini 3 Flash (low reas.)	—	.784	.271	.593	.583	.654
Q2L-bal (ours)	241M	.594	.420	.413	.428	.771
Q2L-agg (ours)	241M	.457	.409	.449	.431	.772
Claude Opus 4.6	—	.658	.082	.327	.404	.598
GLM-4.6V (reas.)	10.3B	.492	.123	.291	.364	.563

Table 6: SenBen results on 2,000 test frames.

\mathrm{R}_{\mathrm{SB}}

\text{F1}_{\mathrm{SB}}

: SenBen Recall/F1,

\text{F1}^{\mathrm{tag}}

: tag F1,

R^{\mathrm{obj}}

: obj. recall, cap: caption sim.

Model	Params	Tags	$\text{F1}^{\mathrm{tag}}$	$\text{F1}^{\mathrm{s}}$
Q2L-bal (ours)	241M	16/16	.594	.847
Q2L-agg (ours)	241M	16/16	.457	.835
Azure Content Safety	—	5/16	.430	.504
OpenAI Moderation	—	6/16	.411	.664
LlavaGuard 1.2	7.0B	6/16	.384	.583
Google SafeSearch	—	8/16	.341	.476

Table 7: Tag detection comparison. Tags:covered MECD tags.

\text{F1}^{\mathrm{tag}}

: macro tag F1 over supported tags.

\text{F1}^{\mathrm{s}}

: safe/unsafe F1.

Panel of Figure 3 — Figure 3: Qualitative SenBen annotations from our custom annotation web app. Each frame shows bounding boxes, predicate arrows, and the object panel with attributes. The scene graphs capture diverse sensitive content: single-object violence (a), multi-actor violence (b), and substance use (c), all using the canonical SenBen vocabulary.

Model	Params	ms/fr	VRAM	$/2K	$\text{F1}_{\mathrm{SB}}$
Q2L-bal (ours)	241M	733	1.2 GB	0	.428
Claude Sonnet 4.6	—	3438	—	12.14	.339
Claude Opus 4.6	—	4555	—	20.02	.404
Gemini 3 Pro † (low reas.)	—	5579	—	26.58	.647
Qwen3-VL-8B	8.3B	5,614	18.8 GB	0	.340
Gemini 3 Flash (low reas.)	—	6121	—	5.80	.583

Table 8: Inference efficiency. Latency:sequential 5-frame avg. VRAM:peak GPU memory. $/2K:total API cost for 2000 frames.

Content Moderation

Content moderation studies how to detect, interpret, and respond to harmful or policy-violating material across digital platforms.

See all 82 papers

SceneGraphVLM: Dynamic Scene Graph Generation from Video with Vision-Language Models
May 13, 2026Vladislav Makarov, Mark Gizetdinov, Dmitry YudinScene Graph Generation Vision-Language Models
MIRAI · MIREA · MIRAI, AXXX
RuleSafe-VL: Evaluating Rule-Conditioned Decision Reasoning in Vision-Language Content Moderation
May 8, 2026Zhifeng Lu, Dianyuan Wang, Yuhu Shang +1VLM Evaluation Logical Reasoning
Beijing University of Posts and Telecommunications, Beijing, China
ZeroBench: An Impossible Visual Benchmark for Contemporary Large Multimodal Models
Feb 13, 2025Jonathan Roberts, Mohammad Reza Taesiri, Ansh Sharma +31Visual Reasoning VLM Reasoning
University of Cambridge

May 13, 2026J/K move · Enter open · S save

SceneGraphVLM: Dynamic Scene Graph Generation from Video with Vision-Language Models

Vladislav Makarov, Mark Gizetdinov, Dmitry Yudin

MIRAI · MIREA · MIRAI, AXXX

Scene Graph Generation Vision-Language Models Video SGG Scene Graphs

Open paper AskPDF arXiv

Scene graph generation provides a compact structured representation for visual perception, but accurate and fast graph prediction from images and videos remains challenging. Recent VLM-based methods can generate scene graphs end-to-end as structured text, yet often produce long outputs with irrelevant objects and relations. We present SceneGraphVLM, a compact method for image and video scene graph generation with small visual language models. SceneGraphVLM serializes graphs in a token-efficient TOON format and trains the model in two stages: supervised fine-tuning followed by reinforcement learning with hallucination-aware rewards that balance relation coverage and precision while penalizing unsupported objects and relations. For videos, the model can optionally condition each frame on the previously generated graph, providing lightweight short-term context without tracking or post-processing. We evaluate SceneGraphVLM on PSG, PVSG, and Action Genome. With compact VLMs and vLLM-accelerated decoding, SceneGraphVLM achieves a strong quality-speed trade-off, improves precision-oriented SGG metrics while preserving reasonable recall, and generates complete scene graphs with approximately one-second latency. Code and implementation details are available at: https://github.com/markus0440/SceneGraphVLM.git.

Abstract

Content Moderation

Explore similar work

SceneGraphVLM: Dynamic Scene Graph Generation from Video with Vision-Language Models

RuleSafe-VL: Evaluating Rule-Conditioned Decision Reasoning in Vision-Language Content Moderation

ZeroBench: An Impossible Visual Benchmark for Contemporary Large Multimodal Models