Skip to content
GL3D (Geometric Learning with 3D Reconstruction): a large-scale database created for 3D reconstruction and geometry-related learning problems
Branch: master
Clone or download
Type Name Latest commit message Commit time
Failed to load latest commit information.

GL3D: Geometric Learning with 3D Reconstruction

Example sequence


GL3D (Geometric Learning with 3D Reconstruction) is a large-scale database created for 3D reconstruction and geometry-related learning problems. Most images contained are captured by drones from multiple scales and perspectives with large geometric overlaps, covering urban, rural area, or scenic spots. It also includes small object reconstructions to enrich the data diversity. If you find this dataset useful for your research, please cite:

    author={Shen, Tianwei and Luo, Zixin and Zhou, Lei and Zhang, Runze and Zhu, Siyu and Fang, Tian and Quan, Long},
    title={Matchable Image Retrieval by Learning from Surface Reconstruction},
    booktitle={The Asian Conference on Computer Vision (ACCV},

If you have used the correspondence labels, please also cite:

    title={Geodesc: Learning local descriptors by integrating geometry constraints},
    author={Luo, Zixin and Shen, Tianwei and Zhou, Lei and Zhu, Siyu and Zhang, Runze and Yao, Yao and Fang, Tian and Quan, Long},
    booktitle={European Conference on Computer Vision (ECCV)},

Dataset Description

GL3D contains 90,630 high-resolution images regarding 378 different scenes. Each scene data is reconstructed to generate a triangular mesh model by the state-of-the-art 3D reconstruction pipeline. Refer to [1] for details. For each scene data, we provide the complete image sequence, geometric labels and reconstruction results.


Research works below are supported by GL3D:

Task Reference
Image retrieval MIRorR, ACCV'18
Local descriptor GeoDesc, ECCV'18
Local descriptor ContextDesc, CVPR'19


For image retrieval task, use 224x224 images and refer to MIRorR.

For learning local descriptor, use 1000x1000 images and refer to GeoDesc.

Sources Data Name Chunk Start Chunk End Descriptions
GL3D gl3d_full_size TBA TBA Full-size images of GL3D
GL3D gl3d_224 0 6 224x224 images of GL3D
GL3D gl3d_1000 0 91 1000x1000 images of GL3D

Use script to download the tar files, by passing augments

bash <data_name> <chunk_start> <chunk_end>

For example, to download GL3D 224x224 images, run

bash gl3d_224 0 6 

To extract the files, run

cat download_data_gl3d_224/*.tar.* | tar -xvf -

Dataset Format

 └── <pid> 
       ├── images/*
       ├── geolabel/*
       ├── img_kpts/*.bin
       └── image_list.txt
File Name Data Name Chunk Start Chunk End Task Descriptions
geolabel/cameras.txt gl3d_cams 0 0 Common Camera intrisic/extrinsic parameters, recovered by SfM.
img_kpts/<img_idx>.bin gl3d_kpts 0 58 Common Image keypoints detected by SIFT.
geolabel/corr.bin gl3d_corr 0 7 Local descriptor Image correspondences that haved survived from SfM.
geolabel/mask.bin gl3d_mask 0 8 Image retrieval Overlap masks of image pairs, computed from mesh re-projections.
geolabel/mesh_overlap.txt gl3d_mo 0 0 Image retrieval Mesh overlap ratio of image pairs, computed from mesh re-projections.
geolabel/common_track.txt gl3d_ct 0 0 Image retrieval Common track ratio of image pairs, computed from SfM.

Again, use to fetch the above geometric labels or reconstruction results,

For data organization, refer to docs/

Python-based IO utilities are provided to parse the data, refer to utils/

Visualizations and examples of usage can be found in example/

Please feel free to inform us if you need some other intermediate results for your research.

Data Preview

The mesh reconstruction is available for preview by substituting <pid> in the following link:<pid>

An example is provided here. Noted that some projects are not online available, from 000000000000000000000000 to 00000000000000000000001d.


This dataset is prepared and maintained by Zixin Luo, Tianwei Shen, Jacky Tang and Tian Fang. 3D reconstructions are obtained by Altizure.

You can’t perform that action at this time.