Yuto Kanda

Yuto Kanda, Hayato Tanoue, Takayuki Hori

CASTLE 2026 asks 185 multiple-choice questions over 600+ hours of synchronized multi-view egocentric video. We explore two approaches on top of a shared multimodal preprocessing layer, including per-person timelines, speaker-resolved transcripts, and multi-VLM caption ensembles. Approach A, SVA: Search-Verify-Answer, is a three-stage pipeline that hierarchically narrows to a primary window, verifies sub-windows with a VLM under four anti-confabulation rules, and fuses evidence with an LLM judge under an evidence-priority hierarchy. Approach B, TMKG: Temporal-Multimodal-Knowledge-Graph, is the contrast: it builds a temporal multimodal knowledge graph, locates a primary cell via graph search, and produces the final answer with a single grounded VLM. SVA reaches a leaderboard accuracy of 0.50 and is our final challenge submission; TMKG reaches 0.35.

Hayato Tanoue, Hiroki Nishihara, Yuma Suzuki et al.

This report presents the CuriosAI team's submission to the EgoExo4D Proficiency Estimation Challenge at CVPR 2025. We propose two methods for multi-view skill assessment: (1) a multi-task learning framework using Sapiens-2B that jointly predicts proficiency and scenario labels (43.6 % accuracy), and (2) a two-stage pipeline combining zero-shot scenario recognition with view-specific VideoMAE classifiers (47.8 % accuracy). The superior performance of the two-stage approach demonstrates the effectiveness of scenario-conditioned modeling for proficiency estimation.