--- license: apache-2.0 pipeline_tag: any-to-any --- # EgoM2P: Egocentric Multimodal Multitask Pretraining **EgoM2P** (ICCV 2025, EgoVis Distinguished Paper Award) is a large-scale egocentric multimodal and multitask model pretrained on eight egocentric datasets. It handles four modalities — RGB and depth video, gaze dynamics, and camera trajectories — and supports tasks such as monocular egocentric depth estimation, camera tracking, gaze estimation, and conditional egocentric video synthesis. - **Paper:** [EgoM2P: Egocentric Multimodal Multitask Pretraining](https://huggingface.co/papers/2506.07886) - **Project page:** [https://egom2p.github.io](https://egom2p.github.io) - **Code:** [https://github.com/ligengen/EgoM2P](https://github.com/ligengen/EgoM2P) The pretrained model weights and tokenizers are available in this repository. Please refer to the GitHub repository for installation, data preparation, tokenization, training, and inference instructions.