Singapore

Surveillance video captioning aims to generate natural language 
descriptions for security video and plays a crucial role in 
applications such as anomaly detection and behavior analysis. 
While vision-language models (VLMs) perform well on general 
video captioning, they often struggle in surveillance scenarios 
due to domain discrepancies in visual style and content 
emphasis. Moreover, reliable evaluation of caption quality in 
this domain also remains challenging. We propose a 
comprehensive framework that addresses both domain 
adaptation and evaluation. For domain adaptation, we enhance 
VLMs via reinforcement fine-tuning, exploring multiple reward 
designs to optimize relevance and accuracy while helping to
improve out-of-domain generalization. For evaluation, we 
develop an LLM-as-a-Judge process that assesses generated 
descriptions from two complementary perspectives: a local 
semantic score capturing fine-grained action fidelity and a global 
quality score evaluating linguistic fluency and temporal 
coherence. Experimental results show that our domain adaptation 
approach outperforms both zero-shot and supervised fine-tuning
baselines. Ablation studies further validate the effectiveness of 
our reward design, training strategy, and out-of-domain 
generalization. In addition, empirical analyses show that the 
proposed LLM-as-a-Judge framework aligns closely with human 
evaluations and exhibits stable and robust performance across 
different conditions.

AAAI 2026

Domain-Adaptive Video 
Captioning for Surveillance 
Videos

surveillance video captioning; vision-language models; domain adaptation; grpo; reinforcement fine-tuning; llm-as-a-judge; anomaly detection; temporal coherence

ICPR 2020 is the premier international conference on Pattern Recognition, gathering leading researchers from around the world to present their latest achievements in Pattern Analysis and Recognition, applied Artificial Intelligence and Machine Learning. The topic of the year is " Putting Artificial intelligence to work on Patterns". 
For the first time, ICPR has employed a two-round review process similar to journal submissions, with new papers allowed to be submitted in either the first or second round and papers submitted in the first round and not accepted allowed to be revised and re-submitted for second round review. 

In this edition of the **ICPR 2020** Conference, there are 1407 Oral/Poster presentations organized in five Tracks: 
- Artificial Intelligence, Machine Learning for Pattern Analysis; 
- Biometrics, Human Analysis and Behavior Understanding; Computer Vision, Robotics and Intelligent Systems; 
- Document and Media Analysis; and 
- Image and Signal Processing. 

The **ICPR 2020** Main Conference program also includes 7 Keynotes from prominent world-renowned scientists, 8 Tutorials, 28 Demos, 8 Challenges, 2 Panels, as well as an Industry/Sponsor Exhibition. In addition, 41 Workshops on a range of important current and emerging research topics complete the ICPR 2020 program. 


ICPR 2020

workshop paper

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-26 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.<br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to AAAI 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://aaai.org/conference/aaai/aaai-26/registration/) first.

**Online Registration Form**: https://aaai.getregistered.net/conference-2026 

Registration Required

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

Learning additional reward signals---beyond the intrinsic rewards 
from the environment---can guide the agent more quickly toward 
optimal policies during training.
Recent work suggested training autoregressive models to 
maximize the likelihood of states observed in expert trajectories.
States that are likely then render a positive reward.
Such methods closely align with anomaly detection (AD), where 
the task is to find observations that deviate significantly from 
typical behavior.
We frame reward generation as an AD problem, generalizing 
existing techniques and proposing a novel methodology for 
learning reward signals. 
Specifically, we train state-of-the-art AD models capable of 
recognizing normal behavior from expert observations while 
actively learning to detect anomalous behavior from the agent's
sub-optimal observations, thus improving the reward learning 
process. We show that recent autoregressive reward-learning 
methods are, in fact, implementations of basic AD methods. 
Empirically, we show the effectiveness of the proposed reward
function across various tasks.


Anomaly-driven 
Reinforcement Learning

Anomaly detection (AD) for safety-critical IoT time series should be 
judged at the event level—by reliability and earliness under 
realistic perturbations. Yet many studies still emphasize point-level 
results on base form of curated public datasets, which limits their
value for model selection in practice. We introduce a unified 
evaluation protocol that adds event-level augmentation simulating 
real-world problems by (i) adds a calibrated sensor dropout stress, 
linear/log drift, additive noise, and window shifts, and (ii) 
performs sensor-level probing via mask-as-missing zeroing with 
per-channel influence estimation to support root-cause analysis.

We evaluate 14 representative models on five public anomaly 
datasets (SWaT, WADI, SMD, SKAB, TEP) and two industrial 
datasets (steam turbine, nuclear turbogenerator) under unified 
splits and event aggregation. There is no universal winner: 
graph-structured models transfer best under dropout and long 
events, for example on SWaT with additive noise a graph 
autoencoder falls from 0.804 to 0.677 (-16\%) while a 
graph-attention variant goes from 0.759 to 0.680 (-10\%) and 
a hybrid graph attention stays nearly flat at 0.762 to 0.756 
(-0.8\%); density or flow models work well on clean, stationary 
plants yet can be fragile to monotone drift, as log drift collapses 
a flow-based detector on SKAB and nuclear power plant (NPP) 
dataset while on SWaT the drop is mild at 0.795 to 0.783 
(-1.5\%); spectral CNNs lead when periodicity is strong; 
reconstruction autoencoders become competitive after basic 
sensor vetting; predictive or hybrid dynamics help when faults 
break temporal dependencies but remain window-sensitive. 
The same protocol also informs architecture choices, for example 
on SWaT under log perturbations replacing normalizing flows with 
a Gaussian density estimator reduces F1 to about 0.57 at high 
stress versus about 0.75 originally, and fixing the learned DAG 
yields a small clean-set gain of about 0.5 to 1.0 points but 
increases drift sensitivity by roughly eight times.


Benchmarking IoT 
Time-Series AD with Event-
Level Augmentations

Reliable anomaly detection (AD) in photovoltaic (PV) modules is 
critical for maintaining solar energy efficiency. However, the 
development of robust computer vision models for PV inspection is 
limited by the scarcity of large-scale, diverse, and balanced datasets.
In this study, we introduce a spatially conditioned denoising diffusion 
probabilistic model, PV-DDPM, designed to generate anomalous 
electroluminescence (EL) images across four PV cell types: 
multi-crystalline silicon (multi-c-Si), mono-crystalline silicon 
(mono-c-Si), half-cut multi-c-Si, and interdigitated back contact 
(IBC) with dogbone interconnect. 
PV-DDPM enables the controlled synthesis of both multi-defect and 
single-defect scenarios by conditioning on binary masks that 
represent structural features and defect positions.
To the best of our knowledge, this is the first framework that jointly 
models multiple PV cell types while allowing simultaneous 
generation of diverse anomaly types.
We also introduce E-SCDD dataset, an enhanced version of the 
SCDD dataset, comprising 1,000 pixel-wise annotated EL images 
spanning 30 semantic classes, and 1,768 unlabeled synthetic 
samples. Quantitative evaluation shows our generated images 
achieve a Fréchet Inception Distance (FID) of 4.10 and Kernel 
Inception Distance (KID) of 0.0023 ± 0.0007 across all categories. 
Furthermore, compared to training on the original SCDD dataset, 
training the vision–language anomaly detection model AA-CLIP 
on E-SCDD improves pixel-level AUC and average precision by 
1.70 and 8.34 points, respectively. Code and data are publicly 
released to ensure reproducibility.

Bridging the Data Gap: Spatially Conditioned Diffusion Model for Anomaly Generation in Photovoltaic Electroluminescence Images

This article introduces Attuned Design Practice (ADP), a design vocabulary and conceptual framework for producing embodied knowledge at the intersection of artistic and scientific inquiry, to inform the development of computational models that generate body-based interpretations and articulations. The convergence of Robotics, AI, HRI, and HCI in Industry 5.0 demand bold experimentation in artistic-scientific inquiry to interpret embodied knowledge beyond quantifiable data. In ADP, we propose the concept of ‘attunement’ to examine how alternative design vocabularies for embodied knowledge can be produced in reciprocity with the development of computational models that direct and facilitate expression and attunement over optimization or efficiency. Attunement is seen as a tuning fork to incrementally investigate the phenomena, relationships and patterns that inherently exist in tensions in-between tangible-intangible, or palpable-impalpable forms of materiality in human-computer connections, partnerships, and shared experiences. We propose Open Rehearsals, where-in alternative experiential understandings unfold in three areas of expertise: (1) preparing of tools and instruments, (2) performative practice, and (3) composition and choreography and notation methods. Our aim is to reinvent design innovation by cultivating integrative, reflective, and adaptive modes of collaborative design-making, from embodied discovery to embodied transformation.

Attuned Design Practice: towards Open Rehearsals for Embodied Knowledge and Computational Design.

This study investigates the intuitive relationship between manual gestures and vocal syllables, grounded in shared neurobiological and motor systems. Participants were tasked with producing "Synekinian pairs," consisting of vocal syllables paired with instinctive gestures. Through qualitative and quantitative analyses, the findings reveal consistent patterns in manual-vocal associations, offering insights into gesture-based sound design, vocal synthesis, gesture notation and human-computer interaction. Future work aims to refine the experimental methodology, expand datasets, and validate these observations across diverse cultural and linguistic contexts.

Exploring “Synekinian Pairs”: Manual-Vocal Gesture Integration in Experimental Contexts

This study investigates how audiences perceive humor in both Human-created and AI-generated Manzai scripts, a traditional Japanese form of scripted stand-up comedy performed by a pair of comedians. We analyze laughter responses using two complementary modalities: explicit laughter, captured through real-time reaction-button inputs, and implicit laughter, measured via mouth-corner motion tracking. Comparative experiments are conducted to evaluate differences between Human-created and AI-generated Manzai scripts. The results reveal that while AI-generated Manzai scripts can elicit explicit laughter to a certain extent, it is less effective in inducing implicit laughter compared to Human-created script. Implicit laughter is observed only when explicit laughter occurred, suggesting that unconscious emotional reactions may precede conscious recognition of humor. Moreover, Human-created Manzai scripts that are contextually relatable and easy to imagine produced larger and more sustained mouth-corner movements, indicating that empathy and contextual understanding play key roles in embodied humor perception. These findings demonstrate that mouth-corner motion is a reliable indicator of subtle amusement and provide new insights into Bodily Expressed Emotion Understanding (BEEU) for AI-generated humor.

Understanding Bodily Expressed Laughter Responses to Human and AI-Generated Japanese Manzai scripts

Predicting Valence-Arousal-Dominance (VAD) dimensions from bodily-expressed emotions in videos remains a fundamentally challenging task in affective computing, requiring models that capture subtle spatiotemporal patterns while balancing computational efficiency and interpretability. We present a comprehensive investigation of VAD prediction approaches on the newly introduced Annotated Bodily Expressed Emotion (ABEE) dataset, which contains approximately 3,200 video clips spanning 8 primary emotion categories and 20 subcategories. We explore two complementary methodologies: a feature-based gradient boosting approach using XGBoost with carefully engineered spatiotemporal features and dimensionality reduction, and deep learning architectures capable of learning hierarchical representations directly from raw video data. Our feature-based approach demonstrates exceptional computational efficiency, with sub-second training times and minimal resource requirements, while our deep models reveal the fundamental difficulty of capturing continuous VAD dimensions from bodily expressions. Through systematic evaluation on the ABEE dataset, we establish baseline performance for the VAD prediction task, achieving $R^2$ scores of -0.090, -0.014, and -0.058 for valence, arousal, and dominance, respectively, with our gradient boosting approach. These results highlight the substantial gap between current methodologies and the inherent complexity of bodily emotion signals, providing benchmarks for future research. We further discuss critical insights regarding feature engineering, temporal dynamics, and the intrinsic challenges of continuous emotion prediction from naturalistic video data, emphasizing the need for dedicated spatiotemporal modeling strategies tailored to bodily expressions.

Spatiotemporal Modeling of Bodily Emotional Expressions for Continuous Valence-Arousal-Dominance Prediction in Video

Sketching is a widely used medium for generating and exploring early-stage design concepts. While generative AI (GenAI) chatbots are increasingly used for idea generation, designers often struggle to craft effective prompts and find it difficult to express evolving visual concepts through text alone. In the formative study (N=6), we examined how designers use GenAI during ideation, revealing that text-based prompting disrupts creative flow. To address these issues, we developed TalkSketch, an embedded multimodal AI sketching system that integrates freehand drawing with real-time speech input. TalkSketch aims to support a more fluid ideation process through capturing verbal descriptions during sketching and generating context-aware AI responses. Our work highlights the potential of GenAI tools to engage the design process itself rather than focusing on output.

TalkSketch: Multimodal Generative AI for Real-time Sketch Ideation with Speech

The capacity for complex, evidence-grounded, and strategically adaptive persuasion remains a formidable grand challenge for artificial intelligence. Prior work, like IBM Project Debater, focused on generating isolated persuasive speeches in highly simplified and shortened debate formats for lay audiences. We introduce a novel autonomous system capable of participating in and winning a full, unmodified two-team competitive policy debate. Our system employs a hierarchi cal architecture of specialized multi-agent workflows, where teams of LLM-powered agents collaborate and critique one an- other to perform discrete argumentative tasks. Each workflow
utilizes iterative retrieval, synthesis, and self-correction using a massive corpus of policy debate evidence (OpenDebateEvidence) [Roush et al. 2024] and produces complete speech transcripts, cross-examinations, and rebuttals. We introduce a live, interactive end-to-end presentation pipeline that renders debates with AI speech and animation: transcripts are surface realized and synthesized to audio with OpenAI text-to-speech (gpt-4.1-tts), and then displayed as talking-head portrait
videos with EchoMimic V1 [Chen et al. 2024, 2025, OpenAI 2025e,c]. Beyond fully autonomous (AI vs. AI) matches, the system supports hybrid human–AI operation: human debaters can intervene at any stage, and humans can optionally serve as opponents against the AI in any speech, enabling AI–human as well as AI–AI rounds. In preliminary evaluations against human-authored cases, our system produces qualitatively superior argumentative components and consistently wins simulated rounds as adjudicated by an independent autonomous judge. Expert human debate coaches also prefer the arguments, evidence, and cases constructed by our system.

A superpersuasive autonomous policy debating system

Despite remarkable advances in machine learning, real-time, online inference of agents’ intentions remains one of the enduring grand challenges in AI. Plan, activity, intent, and goal recognition involve making inferences about actors from observations of their behavior, whether they are software agents, robots, or humans. This tutorial targets AI students and researchers interested in exploring these recognition tasks to enhance their ongoing research or identify new research directions.
Attendees will leave with a clear understanding of core concepts, tasks, and relationships among plan, activity, intent, and goal recognition, as well as exposure to the state-of-the-art computational representations, algorithms, and empirical methodologies. The tutorial is innovative in its unification of fragmented research across multiple AI subfields and its use of live demonstrations to connect theory with practice. Building on the successful 2019 AAAI tutorial on the same topic (Keren, Mirsky, and Geib 2019), this session offers an updated and expanded perspective. It incorporates new material from the presenters’ co-authored book (Mirsky, Keren, and Geib 2021), along with the latest developments and advances in the field. Prerequisite knowledge: Familiarity with general AI concepts, probabilistic reasoning, and planning is recommended. Basic knowledge of Bayesian inference and logic-based reasoning will be helpful but not required.

Downloads

Next from AAAI 2026

Anomaly-driven Reinforcement Learning

Similar lecture

Temporal Feature Enhancement Network with External Memory for Object Detection in Surveillance Video

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES