GV-Bench

GV-Bench: Benchmarking Local Feature Matching for Geometric Verification of Long-term Loop Closure Detection
Jingwen Yu, Hanjing Ye, Jianhao Jiao, Ping Tan, Hong Zhang
2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
arXiv, IEEEXplore, Project Page, Blog Post (in Chinese), Video (Bilibili)

Abstract

In short, GV-Bench provides a benchmark for evaluating different local feature matching methods on geometric verification (GV), which is crucial for vision-based localization and mapping systems (e.g., Visual SLAM, Structure-from-Motion, Visual Localization).

News

• ⭐ Add support for Image-matching-models, thanks for their excellent work! 😄
• 🎉 Chinese intro on wechat offical account.
• ⭐ Paper is release on arxiv.
• 🎉 The paper is accepted by IROS 2024!
• 🚀 Releasing the visualization of image matching results.
• 🚀 Releasing the benchmark!

Installation

# clone the repo
git clone --recursive https://github.com/jarvisyjw/GV-Bench.git
conda create -n gvbench python=3.11
pip install -r requirements
cd third_party/image-matching-models
git submodule init
git submodule update

Evaluation

Data

1. Get the GV-Bench sequences from here.
2. Unzip and organize the dataset folder like following:

  |-- gt
  |   |-- day.txt
  |   |-- night.txt
  |   |-- nordland.txt
  |   |-- season.txt
  |   |-- uacampus.txt
  |   |-- weather.txt
  |-- images
  |   |-- day0
  |   |-- day1
  |   |-- night0
  |   |...

Benchmark

Now, we support using image-matching-models directly, which enables many more state-of-the-art matching methods for Geometric Verification (GV). The example usage:

# Show all supported image matching models
python main.py --support_model
# Run
python main.py config/day.yaml

In the configs, please specify the data directory, sequence info (which is provided with the repo in dataset/gt folder), image_size and the matcher you want to use. The evaluation supports runing multiple matchers in a single run. However, our sequences provides a rather large number of pairs, so the evaluation might takes time.

If you want to replicate the paper's result of IROS2024, please refer to this.

Using Customized Image Matcher

• We recommend contributing or building your image matcher as the standard of image-matching-models.
• Example usage of bench loader and evaluator

  # bench sequence
  gvbench_seq = ImagePairDataset(config.data, transform=None) # load images
  labels = gvbench_seq.label # load labels
  MODEL = YOUR_MODEL(max_num_keypoints=2048) # your image matching model
  # if your method is two-stage image matching
  # i.e. Step 1: Keypoints Extraction
  #      Step 2: Keypoints matching
  # We recommend you set a max_num_keypoints to 2048
  # to be consistent with the image-matching-models default setup.
  scores = []
  for data in gvbench_seq:
    img0 = load_image(data['img0'], resize)
    img1 = load_image(data['img1'], resize)
    inliers = MODEL(img0, img1)
    scores.append(inliers)
  # normalize
  scores_norm = (scores - np.min(scores)) / (np.max(scores)- np.min(scores))
  mAP, MaxR = eval(scores, labels)

Experiments

Seq: Day
+---------------------------------------------------------+                                                
|                      GV-Bench:day                       | 
+---------------+--------------------+--------------------+
|    Matcher    |        mAP         |   Max Recall@1.0   |
+---------------+--------------------+--------------------+
|     master    | 0.9975692177820614 | 0.5236404833836859 |
| superpoint-lg | 0.9937596599094175 | 0.4628776435045317 |
|    disk-lg    | 0.9971895284342447 | 0.5448640483383685 |
|      roma     | 0.9961146810617404 | 0.389916918429003  |
|    sift-nn    | 0.9742650091865308 | 0.3534365558912387 |
|    sift-lg    | 0.9934969495417851 | 0.4178247734138973 |
|     duster    | 0.9913920054359513 | 0.5118202416918429 |
|     xfeat     | 0.9896198466632193 | 0.4217522658610272 |
|     loftr     | 0.9933317967081768 | 0.41774924471299096|
+---------------+--------------------+--------------------+

Seq: Night
+----------------------------------------------------------+                                                
|                      GV-Bench:night                      |                                                  
+---------------+--------------------+---------------------+
|    Matcher    |        mAP         |    Max Recall@1.0   |
+---------------+--------------------+---------------------+
|     master    | 0.9868066941492244 |  0.504996776273372  |
| superpoint-lg | 0.9767695728913396 |  0.384107027724049  |
|    disk-lg    |  0.96613211965037  | 0.42109929078014185 |
|      roma     | 0.9878749584388696 | 0.11210509348807221 |
|    sift-nn    | 0.6135980930535652 | 0.04223081882656351 |
|    sift-lg    | 0.9343053648298021 | 0.45970341715022567 |
|     duster    | 0.7884898363259103 | 0.24186009026434557 |
|     xfeat     | 0.9142595500480328 |  0.2730496453900709 |
|     loftr     | 0.9849044142298791 | 0.3534816247582205  |
+---------------+--------------------+---------------------+

Seq: Season
+-----------------------------------------------------------+                                                
|                      GV-Bench:season                      |                                                
+---------------+--------------------+----------------------+                                                
|    Matcher    |        mAP         |    Max Recall@1.0    |                                                
+---------------+--------------------+----------------------+                                                
|     master    | 0.9988580197550959 |  0.6730315734311542  |                                                
| superpoint-lg | 0.998920429558958  |  0.7452181317961483  |                                                
|    disk-lg    | 0.999232632789443  |  0.7445958338792087  |                                                
|      roma     | 0.9980641689106753 | 0.09049521813179615  |                                                
|    sift-nn    | 0.9791038717843851 | 0.24914843442945106  |                                                
|    sift-lg    | 0.9985371274637678 |  0.5848617843573956  |                                                
|     duster    | 0.9626673775950897 | 0.030558102973928993 |                                                
|     xfeat     | 0.998104560026835  |  0.7191143718066291  |                                                
|     loftr     | 0.9989917476116729 |  0.7945106773221539  |                                                
+---------------+--------------------+----------------------+ 

Seq: Weather
+----------------------------------------------------------+
|                     GV-Bench:weather                     |
+---------------+--------------------+---------------------+
|    Matcher    |        mAP         |    Max Recall@1.0   |
+---------------+--------------------+---------------------+
|     master    | 0.9988330039057637 |  0.4650407500459587 |
| superpoint-lg | 0.9984747637720905 |  0.566486917090508  |
|    disk-lg    | 0.9984560239543505 |  0.5313131932103683 |
|      roma     | 0.9973693543564228 | 0.11639806360683866 |
|    sift-nn    | 0.9962779437782742 |  0.416355168821619  |
|    sift-lg    | 0.9983667779883508 |  0.5204363012439488 |
|     duster    | 0.9984571842125228 | 0.23999632330412402 |
|     xfeat     | 0.9973036794158792 | 0.38568539738954594 |
+---------------+--------------------+---------------------+

Seq: UAcampus
+-----------------------------------------------------------+                                                
|                      GV-Bench:uacampus                    | 
+---------------+--------------------+----------------------+
|    Matcher    |        mAP         |    Max Recall@1.0    |
+---------------+--------------------+----------------------+
|     master    | 0.8295730669487469 | 0.12686567164179105  |
| superpoint-lg | 0.8518262736710289 | 0.26119402985074625  |
|    disk-lg    | 0.7876584656153724 | 0.06716417910447761  |
|      roma     | 0.7335515971553305 | 0.06716417910447761  |
|    sift-nn    | 0.5002679213167198 | 0.007462686567164179 |
|    sift-lg    | 0.814838399559011  | 0.22885572139303484  |
|     duster    | 0.5915022125735299 | 0.022388059701492536 |
|     xfeat     | 0.8190512921138375 | 0.19402985074626866  |
|     loftr     | 0.7951256302225166 |  0.2263681592039801  |
+---------------+--------------------+----------------------+

Seq: Nordland
+-------------------------------------------------------------+                                                
|                      GV-Bench:nordland                      | 
+---------------+---------------------+-----------------------+
|    Matcher    |         mAP         |     Max Recall@1.0    |
+---------------+---------------------+-----------------------+
|     master    |  0.8351055461236119 |   0.1266891891891892  |
| superpoint-lg |  0.7352313891322164 |  0.02027027027027027  |
|    disk-lg    |  0.7109283125369613 |  0.11148648648648649  |
|      roma     |   0.77038309764354  |  0.010135135135135136 |
|    sift-nn    | 0.21872420539646553 | 0.0016891891891891893 |
|    sift-lg    |  0.7351407725535801 |  0.06418918918918919  |
|     duster    | 0.15624893409492682 |  0.012387387387387387 |
|     xfeat     |  0.7170417005176157 |  0.14864864864864866  |
|     loftr     |  0.7842976503985124 |  0.14695945945945946  |
+---------------+---------------------+-----------------------+

Acknowledgement

• This work thanks hloc, image-matching-models for their amazing works.
• Contact: jingwen.yu@connect.ust.hk

Citation

@inproceedings{yu2024gv,
  title={GV-Bench: Benchmarking Local Feature Matching for Geometric Verification of Long-term Loop Closure Detection},
  author={Yu, Jingwen and Ye, Hanjing and Jiao, Jianhao and Tan, Ping and Zhang, Hong},
  booktitle={2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={7922--7928},
  year={2024},
  organization={IEEE}
}