Home

Awesome

<div style="text-align: center;"> <h2>[ECCV2024] Diff-Reg: Diffusion Model in Doubly Stochastic Matrix Space for Registration Problem [<a href="https://arxiv.org/pdf/2403.19919">Arxiv</a>|<a href="https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08193.pdf">ECCV2024</a>]</h2> </div>

👀 If you have any questions, please let me (wuqianliang@njust.edu.cn) know~

Installation

1. Please use the NVIDIA TITAN RTX or NVIDIA GeForce RTX 3090 GPU !! **If you switch to an RTX 4090 or a higher version GPU, you will need to re-train the model. We have test the Diff-Reg-4dmatch on the RTX 4090 GPU.**

2. Please utilize commands 'conda env create -f Diff-Reg-2d3d/eccv24_2d3d_env.yml', 'conda env create -f Diff-Reg-3dmatch/eccv24_3d_env.yml', and 'conda env create -f Diff-Reg-4dmatch/eccv24_4d_env.yml' to install environments for three tasks.

Pre-trained Weights

Please look at the release page for the pre-trained model weights of three experiments.

Data Preparation && Training

Our 2D3D registration code is mainly based on 2D3D-MATR, and our 3d registration code is based on Lepard. Please refer to Lepard and 2D3D-MATR.

Inference

For 3DMatch and 4DMatch:

python main.py --config configs/test/3dmatch.yaml --test_epoch=$epoch

python main.py --config configs/test/4dmatch.yaml --thr=0.55 --test_epoch=$epoch

For 2D-3D registration:

cd experiments/2d3dmatr.rgbdv2.stage4.level3.stage1; sh eval.sh $epoch

Results

Quantitative results on the 4DMatch and 4DLoMatch benchmarks. The best results are highlighted in bold, and the second-best results are underlined.

CategoryMethod4DMatch NFMR↑4DMatch IR↑4DLoMatch NFMR↑4DLoMatch IR↑
Scene FlowPointPWC21.6020.010.07.20
FLOT27.1024.9015.2010.70
Feature MatchingD3Feat55.5054.7027.4021.50
Predator56.4060.4032.1027.50
Lepard83.6082.6466.6355.55
GeoTR83.2082.2065.4063.60
RoITr83.0084.4069.4067.60
DDPMDiff-Reg (Backbone)85.4781.1572.3759.50
Diff-Reg (steps=1)85.2383.8573.1965.26
Diff-Reg (steps=20)90.2587.9877.1567.00

Quantitative results on the 3DMatch and 3DLoMatch benchmarks. The best results are highlighted in bold, and the second-best results are underlined.

MethodReference3DMatch FMR↑3DMatch IR↑3DMatch RR↑3DLoMatch FMR↑3DLoMatch IR↑3DLoMatch RR↑
FCGFICCV 201995.2056.9088.2060.9021.4045.80
D3FeatCVPR 202095.8039.0085.8069.3013.2040.20
PredatorCVPR 202196.7058.0091.8078.6026.7062.40
LepardCVPR 202297.9557.6193.9084.2227.8370.63
GeoTRCVPR 202298.1072.7092.3088.7044.7075.40
RoITrCVPR 202398.0082.6091.9089.6054.3074.80
PEAL-3DCVPR 202398.5073.3094.2087.6049.0079.00
Diff-RegECCV 202496.2830.9295.0069.609.6073.80

Evaluation results on RGB-D Scenes V2 [Li et al. 2023]. The best results are highlighted in bold, and the second-best results are underlined.

ModelScene-11Scene-12Scene-13Scene-14Mean
Mean depth (m)1.741.661.181.391.49
Inlier Ratio ↑
FCGF-2D3D6.88.511.85.48.1
P2-Net9.712.817.09.312.2
Predator-2D3D17.719.417.28.415.7
2D3D-MATR32.834.439.223.332.4
FreeReg36.634.534.218.230.9
Diff-Reg (dino)38.637.445.431.638.3
Diff-Reg (dino/backbone)44.949.538.333.141.4
Diff-Reg (dino/steps=1)47.548.932.822.437.9
Diff-Reg (dino/steps=10)47.248.732.922.437.8
Feature Matching Recall ↑
FCGF-2D3D11.1030.4051.5015.5027.10
P2-Net48.6065.7082.5041.659.60
Predator-2D3D86.1089.2063.9024.3065.90
2D3D-MATR98.6098.0088.7077.9090.80
FreeReg91.9093.4093.1049.6082.00
Diff-Reg (dino)100.0100.089.7081.992.9
Diff-Reg (dino/backbone)100.0100.092.891.296.0
Diff-Reg (dino/steps=1)100.0100.088.776.591.3
Diff-Reg (dino/steps=10)100.0100.088.777.091.4
Registration Recall ↑
FCGF-2D3D26.441.237.116.830.4
P2-Net40.340.241.231.938.4
Predator-2D3D44.441.221.613.730.2
2D3D-MATR63.953.958.849.156.4
FreeReg+Kabsch38.751.630.715.534.1
FreeReg+PnP74.272.554.527.957.3
Diff-Reg (dino)87.586.363.960.674.6
Diff-Reg (dino/backbone)79.286.375.371.278.0
Diff-Reg (dino/steps=1)98.6100.087.666.888.3
Diff-Reg (dino/steps=10)98.696.183.563.785.5

Evaluation results on 7Scenes [Li et al. 2023]. The best results are highlighted in bold, and the second-best results are underlined.

ModelChessFireHeadsOfficePumpkinKitchenStairsMean
Mean depth (m)1.781.550.802.032.252.131.841.77
Inlier Ratio ↑
FCGF-2D3D [Choy et al. 2019]34.232.814.826.023.322.56.022.8
P2-Net [Choy et al. 2019]55.246.713.036.232.032.85.831.7
Predator-2D3D [Huang et al. 2021]34.733.816.625.923.122.27.523.4
2D3D-MATR [Li et al. 2023]72.166.031.360.750.252.518.150.1
Diff-PnP (dino/backbone)79.271.054.170.455.860.222.959.1
Diff-PnP (dino/steps=10)73.360.845.563.147.853.320.452.0
Feature Matching Recall ↑
FCGF-2D3D [Choy et al. 2019]99.798.269.997.183.087.716.278.8
P2-Net [Choy et al. 2019]100.099.358.999.187.292.216.279.0
Predator-2D3D [Huang et al. 2021]91.395.176.788.679.280.631.177.5
2D3D-MATR [Li et al. 2023]100.099.698.6100.092.495.958.192.1
Diff-PnP (dino/backbone)100.0100.0100.0100.091.398.158.192.5
Diff-PnP (dino/steps=10)100.098.597.3100.087.896.860.891.6
Registration Recall ↑
FCGF-2D3D [Choy et al. 2019]89.579.719.285.969.479.06.861.4
P2-Net [Choy et al. 2019]96.986.520.591.775.385.24.165.7
Predator-2D3D [Huang et al. 2021]69.660.717.862.956.262.69.548.5
2D3D-MATR [Li et al. 2023]96.990.752.195.580.986.128.475.8
Diff-PnP (dino/backbone)100.094.090.499.381.294.627.083.8
Diff-PnP (dino/steps=10)99.394.391.899.179.991.825.783.1

:hearts: Acknowledgement

We thank the respective authors of Lepard,2D3D-MATR, GeoTR,RoITR,GraphSCNet, and Vision3D for their open source code.

Citation

Please consider citing the following BibTeX entry if you find our work helpful for your research.

@article{wu2024diff,
  title={Diff-Reg v1: Diffusion Matching Model for Registration Problem},
  author={Wu, Qianliang and Jiang, Haobo and Luo, Lei and Li, Jun and Ding, Yaqing and Xie, Jin and Yang, Jian},
  journal={arXiv preprint arXiv:2403.19919},
  year={2024}
}