CVPR2024 - MCNet: Rethinking the Core Ingredients for Accurate and Efficient Homography Estimation

Official implementation of MCNet from the paper:
"MCNet: Rethinking the Core Ingredients for Accurate and Efficient Homography Estimation"
(CVPR 2024)

📌 Table of Contents

• 📌 Table of Contents
• 🚀 Introduction
• 📦 Environment Setup
• 💾 Dataset Preparation
• 💻 Quick Start
• 📊 Experimental Results
• 🔍 Visualization
• 📚 Citation
• 📧 Contact

---

🚀 Introduction

MCNet is a multiscale correlation searching network designed for efficient and accurate homography estimation. It achieves state-of-the-art performance with significant reductions in computational cost compared to previous methods. Key innovations include:

1. Multiscale Correlation Searching: Combines multi-scale strategies with iterative correlation searching to improve accuracy with minimal resource growth.
2. Fine-Grained Optimization (FGO) Loss: Dynamically adjusts training gradients during convergence, improving precision without extra overhead.
3. Cross-Modality Evaluation: Evaluation across different datasets and modalities (GoogleEarth, GoogleMap, MSCOCO, SPID) shows the effectiveness of the network in different scenarios.

---

📦 Environment Setup

Our implementation has been successfully tested on NVIDIA RTX4090/3090/2080/8000 GPU platfroms. Recommended setup using Anaconda:

# Create and activate environment
conda create -n mcnet python=3.8
conda activate mcnet

# Install dependencies
# refer to https://pytorch.org/get-started/previous-versions/
pip install torch==1.10.1+cu111 torchvision==0.11.2+cu111 -f https://download.pytorch.org/whl/cu111/torch_stable.html
pip install numpy opencv-python matplotlib cupy==10.6 

---

💾 Dataset Preparation

Supported datasets: MSCOCO, GoogleEarth, GoogleMap, SPID and any other personal dataset. Follow these steps:

1. Download datasets.
2. Configure dataset paths in datasets.py:

---

💻 Quick Start

Training

python train.py --gpuid 0 --dataset mscoco --batch_size 16

Testing

python test.py \
    --gpuid 0 \
    --dataset mscoco \
    --checkpoint ./weights/mscoco.pth

Pretrained Weights

Dataset	Weight File
MSCOCO	weights/mscoco.pth
GoogleEarth	weights/ggearth.pth
GoogleMap	weights/ggmap.pth
SPID	weights/spid.pth

---

📊 Experimental Results

Main Metrics Comparison

Method	MACE ↓	Inference Time(ms) ↓	Params(M) ↓
IHN (2-scale)	0.060	60.1	1.71
RHWF (2-scale)	0.039	157.1	1.29
MCNet	0.031	33.1	0.85

---

🔍 Visualization

MACE Comparison

📚 Citation

If this work helps your research, please cite:

@inproceedings{zhu2024mcnet,
  title={Mcnet: Rethinking the core ingredients for accurate and efficient homography estimation},
  author={Zhu, Haokai and Cao, Si-Yuan and Hu, Jianxin and Zuo, Sitong and Yu, Beinan and Ying, Jiacheng and Li, Junwei and Shen, Hui-Liang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={25932--25941},
  year={2024}
}

---

📧 Contact

• Haokai Zhu: hkzhu.zju@gmail.com
• Si-Yuan Cao: cao_siyuan@zju.edu.cn