Know-Show: Benchmarking Video-Language Models on Spatio-Temporal Grounded Reasoning (Under Review) | Aug 2025 – Nov 2025 | Paper | Project Page
Introduced Know-Show, a benchmark for evaluating Spatio-Temporal Grounded Reasoning in Video-Language Models, i.e., assessing a model’s ability to reason about actions and semantics while grounding predictions in spatial and temporal evidence.
Designed a unified evaluation framework integrating reasoning and localization across five scenarios, covering person, object, person–object, hand–object interactions, and temporal dynamics.
Constructed the benchmark from Action Genome and Ego4D, incorporating 2.5K high-quality human-authored questions targeting fine-grained spatial and temporal understanding.
Identified substantial performance gaps between Video-Language Models and humans, particularly in person–object and hand–object co-grounded reasoning.
Proposed GRAM, a training-free plug-in that improves grounded reasoning in Video-Language Models through attention-based video token selection and explicit timestamp encoding with 11% improvement over state-of-the-art Video-Language Models.
Dataset is released here; code will be released upon paper acceptance.
VOST-SGG: VLM-Aided One-Stage Spatio-Temporal Scene Graph Generation (Under Review) | Feb 2025 – Aug 2025 | Paper | Project Page
Addressed two major limitations of DETR-style Spatio-Temporal Scene Graph Generation (ST-SGG) models: uninformed query initialization and reliance on unimodal visual features for predicate classification.
Proposed VOST-SGG, a one-stage ST-SGG framework featuring Dual-Source Query Initialization and a Multi-Modal Feature Bank that integrates visual, textual, and spatial cues derived from Vision-Language Models.
Achieved state-of-the-art performance on the Action Genome benchmark with extensive ablation studies, yielding a 6% performance improvement over prior methods.
Code to be released upon paper acceptance.
IMoRe: Implicit Program-Guided Reasoning for Human Motion Q&A (ICCV 2025) | Aug 2024 – Feb 2025 | Paper | Project Page
Proposed IMoRe, a 3D human motion reasoning framework for Human Motion Question Answering, replacing explicit Neuro-Symbolic program execution with an implicit program-guided neural reasoning approach.
Introduced a program-guided reading mechanism that dynamically selects multi-level motion features from a pre-trained Motion Language Model, capturing both high-level semantics and fine-grained motion cues.
Integrated an iterative memory module that refines evidence across reasoning steps, improving scalability and adaptability across diverse query types.
Achieved state-of-the-art performance on BABEL-QA with a 7% improvement and demonstrated strong cross-dataset generalization on the newly constructed HuMMan-QA benchmark.
Code and HuMMan-QA dataset are released here.
Situational Scene Graph for Structured Human-Centric Situation Understanding (WACV 2025) | Jan 2023 – Aug 2024 | Paper | Project Page
Introduced Situational Scene Graphs (SSGs), a structured video representation that integrates the strengths of Scene Graphs and Situation Frames to represent the current perceptual state of human-centric situations.
Established the novel vision-language task of Situational Scene Graph Generation.
Constructed a high-quality SSG dataset with role-value annotations for persons, objects, and predicates in human–object interactions.
Developed InComNet, demonstrating consistent improvements over state-of-the-art methods in Situational Scene Graph Generation, Situation Recognition, Predicate Classification, and human-centric situation reasoning tasks.
Code and SSG dataset are released here.
Assessing Information Quality of Wikipedia Articles Through Google’s E-A-T Model (IEEE Access 2022) | Paper
Dec 2020 – Dec 2021
Introduced a theoretical evaluation framework grounded in Google’s E-A-T (Expertise, Authoritativeness, and Trustworthiness) model for assessing the quality of Wikipedia articles.
Outperformed existing information-quality assessment approaches, providing a scalable and cost-effective method for evaluating large-scale user-generated content.