CVPR 2025 - MOSE Track of the 4th PVUW challenge

MASSeg : Multi-scale-Aware Segmentation Network for Complex Video Object Segmentation

🏅 Paper on [arXiv] | 🏛️ [ MOSE Homepage ] 

👨‍🚀 Author: Xuqiang Cao, Linnan Zhao, Jiaxuan Zhao, Fang Liu, Puhua Chen, Wenping Ma

Introduce

MASSeg is our 2nd place solution to the CVPR 2025 PVUW MOSE Challenge. It enhances a segmentation framework with mask scaling, MOSE-specific data augmentation, and the MOSE+ dataset, achieving a J&F score of 0.8628.

Installation

MASSeg needs to be installed first before use. The code requires python>=3.10, as well as torch>=2.5.1 and torchvision>=0.20.1. Please follow the instructions here to install both PyTorch and TorchVision dependencies. You can install MASSeg on a GPU machine using:

Clone this repository

git clone --depth 1 https://github.com/cxqNet/MASSeg.git  && cd MASSeg

pip install -e .

Getting Started

Datasets Preparation

We integrated a new dataset "MOSE+" to improve the performance of the model in the training phase The dataset "MOSE+" can be downloaded from this link of BaiduPan. Please place it in the data folder and confirm the file structure as follows:

MASSeg/
    |- data
        |- MOSE+
           |- JPEGImages
           |- Annotations
           |- moseplus.txt
    |- Training
    |- tools
    |- ...

Download Pretrained Weights

We provide the pretrained model weights used in the Track. You can download them via the link below:

🔗 Baidu Netdisk Download Link: [BaiduPan]

After downloading, please place the weights in the following directory:

MASSeg/
    |- checkpoints/
        │- mass_seg_mose.pth

MOSE Challenge Organizers :

Henghui Ding Primary Organizer Fudan University

Chang Liu Primary Organizer Nanyang Technological University

Shuting He Nanyang Technological University

Xudong Jiang Nanyang Technological University

Philip H.S. Torr University of Oxford

Song Bai ByteDance

References

• We especially thank the authors of MOSE for providing a valuable and challenging dataset, which significantly advances research in complex video object segmentation.
• We also appreciate the open-source contributions of Sam2 and Cutie, whose frameworks and ideas greatly inspired and supported our work.