๐Ÿ‘๏ธ EgoCaptioner: Towards Comprehensive and Fine-Grained Egocentric Video Captioning

โœจ Overview

EgoCaptioner describes first-person video with a video-level overview and timestamped steps. The overview establishes the scene and the camera wearer's activity. The steps describe how the activity unfolds, with details on manipulation, locomotion, social interaction, and visual attention.

Built on Qwen3.5-9B, it is trained with coarse-to-fine caption supervision and reinforcement learning for factual content, temporal alignment, and caption style.

EgoCaptioner reward design

๐Ÿš€ Getting Started

Please refer to our GitHub repository for installation and usage.

Downloads last month
15
Safetensors
Model size
9B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Vickyinmyheart824/EgoCaptioner

Finetuned
Qwen/Qwen3.5-9B
Finetuned
(1,447)
this model
Quantizations
2 models