Yegane Bagheri
Graph-Augmented Retrieval for Long-Form Video Question Answering.
Rel. Luca Cagliero. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
|
|
PDF (Tesi_di_laurea)
- Tesi
Accesso limitato a: Solo utenti staff fino al 24 Gennaio 2028 (data di embargo). Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (1MB) |
Abstract
Video RAG systemsLong video question answering is a growing challenge: videos of 30–90 minutes contain rich multimodal content, yet no current vision-language model can process an hour-long video at native resolution. Even state-of-the-art proprietary models like GPT-4o achieve only 48.9% accuracy on LVBench, a benchmark of hour-long video questions where humans score 94.4%. Retrieval-Augmented Generation (RAG) offers a scalable alternative, but existing Video RAG systems treat retrieval as flat text matching, ignoring the relational structure of video content: entity persistence across scenes, temporal ordering of events, and causal dependencies between actions. This thesis presents TC-VideoRAG (Temporal-Causal Video Retrieval-Augmented Generation), a system that augments Video RAG with a heterogeneous graph encoding three types of relationships: knowledge edges (entity relations extracted by REBEL), temporal edges (within-scene and cross-scene event ordering), and causal edges (within-scene via VGCM and cross-scene via a hybrid discovery pipeline).
Each edge type can be independently enabled or disabled, allowing systematic evaluation of which relationships help which question types
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
