[ECCV 26] MSPL: Multi-Step Pseudo-Labeling for Open-Vocabulary Object Detection

Introduction

This is an official release of the paper MSPL: Multi-Step Pseudo-Labeling for Open-Vocabulary Object Detection.

MSPL: Multi-Step Pseudo-Labeling for Open-Vocabulary Object Detection,
Hojun Choi, Youngsun Lim, Jaeyo Shin, Hyunjung Shim [Paper][Project Page][Dataset][Bibtex]

Updates

⛽⛽⛽ Contact: eric970412@gmail.com or hchoi256@kaist.ac.kr

• [2026.06.24] 📄 Our ECCV-version paper is now available on arXiv.
• [2026.06.17] 🎉 MSPL has been accepted to ECCV 2026.
• [2026.04.04] 🎉 MSPL has been accepted to the non-archival CVPR 2026 Workshop MUSI.
• [2026.03.26] The code has been released and will continue to be updated.

Installation

This project builds upon MMDetection 3.x and BARON.

It requires the following OpenMMLab packages:

• MMEngine >= 0.6.0
• MMCV-full >= v2.0.0rc4
• MMDetection >= v3.0.0rc6
• lvisapi

pip install openmim mmengine
mim install "mmcv>=2.0.0rc4"
pip install git+https://github.com/lvis-dataset/lvis-api.git
mim install "mmdet>=3.0.0rc6"
pip install ftfy regex

Quick Start

Obtain CLIP Checkpoints

We use CLIP's ViT-B-16 model for the implementation of our method. pip install git+https://github.com/openai/CLIP.git and run

import clip
import torch
model, _ = clip.load("ViT-B/16")
torch.save(model.state_dict(), 'checkpoints/clip_vitb16.pth')

Offline Multi-Step Pseudo-Labeling

For pseudo-label generation,

• follow the instructions to generate pseudo-labels for base-category training images.
• or directly download the pre-generated pseudo-labels, instances_train2017_pseudo_v0_new_caption.json, from Hugging Face.

Note that OV-COCO and OV-LVIS share the same 118,287 training images, but their annotations differ.

Training and Testing

The training and testing are supported here.

License

This project is released under the NTU S-Lab License 1.0.

Citation

@misc{choi2026msplmultisteppseudolabelingopenvocabulary,
      title={MSPL: Multi-Step Pseudo-Labeling for Open-Vocabulary Object Detection}, 
      author={Hojun Choi and Youngsun Lim and Jaeyo Shin and Hyunjung Shim},
      year={2026},
      eprint={2510.14792},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.14792v4}, 
}