Seeing Culture: A Benchmark for Visual Reasoning and Grounding
A two-stage benchmark where models must first reason about a cultural artifact, then visually ground it, built across Southeast Asia.
Publications
Also on Google Scholar. An asterisk marks equal contribution.
Cultural multimodal reasoning and grounding
Benchmarks and methods that test whether a vision-language model understands a cultural scene or has only learned to sound confident about it. Southeast Asia first, because it is the hardest available test case.
Cultural Moment Benchmark: Evaluating Video Cultural Reasoning and Grounding in Southeast Asia
Extends cultural reasoning and grounding from still images to video, across Southeast Asia. EMNLP 2026.
Retrieval-Augmented Reasoning Segmentation in Cultural Context
Retrieval-augmented reasoning segmentation for culturally grounded scenes, so a model must retrieve the right cultural knowledge before it segments. Under review for ACM TOMM.
Video Object Segmentation in Cultural Context
Tracking and segmenting culturally significant objects through video, where the object’s identity depends on cultural knowledge rather than appearance alone.
Debiased and semantic text-video retrieval
Video retrieval models take shortcuts: clip length, object co-occurrence, verb priors. My doctoral work identified these shortcuts and removed them with semantic role structure and causal intervention.
Towards Debiasing Frame Length Bias in Text-Video Retrieval via Causal Intervention
Shows that text-video retrieval models exploit clip length as a shortcut, and mitigates the bias with causal intervention.
An Overview of Challenges in Egocentric Text-Video Retrieval
An extended abstract mapping open challenges in egocentric text-video retrieval.
RoME: Role-aware Mixture-of-Expert Transformer for Text-to-Video Retrieval
A role-aware mixture-of-experts transformer that separates verbs, objects and context for text-to-video retrieval.
Exploiting Semantic Role Contextualized Video Features for Multi-Instance Text-Video Retrieval (EPIC-KITCHENS-100 Challenge 2022)
Joint 3rd place in the EPIC-Kitchens-100 Multi-Instance Retrieval Challenge at CVPR 2022.
Semantic Role Aware Correlation Transformer for Text to Video Retrieval
Aligns semantic roles between text and video with a correlation transformer for retrieval.
Long-video retrieval and multimodal planning
Grounding language-model plans in video, and finding the right moment inside hours of egocentric footage.
VG-TVP: Multimodal Procedural Planning via Visually Grounded Text-Video Prompting
Procedural planning with LLMs improves when text and video prompts are visually grounded together.
Enhancing Video Corpus Moment Retrieval in Long Ego-centric Videos with LLM and Audio Fusion
Finding the right moment inside hours of egocentric video by fusing audio with language-model reasoning over the visual stream.
Earlier work
Vehicle detection and recognition, from my master's.
Deep Learning Based Vehicle Make-Model Classification
CNN-based vehicle make and model classification from my MSc research.