Skip to content
Matchable Image Retrieval by Learning from Surface Reconstruction (ACCV 2016)
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

MIRorR: Matchable Image Retrieval by Learning from Surface Reconstruction

Mirror is the matchable image retrieval pipeline for 3D reconstruction and related applications. Different from typical object retrieval, matchable image retrieval aims to find similar images with large overlaps. Typical CNN-based methods do not generalize well to this problem because models are trained to find objects of the same category. This project proposes a new method to tackle this problem, by utilizing regional feature aggregation and the accurate auto-annotated 3D geometric data. For more details, you can refer to the paper:

Matchable Image Retrieval by Learning from Surface Reconstruction

Tianwei Shen*, Zixin Luo*, Lei Zhou, Runze Zhang, Siyu Zhu, Tian Fang, Long Quan (* denotes equal contributions)

In ACCV 2018.

Feel free to submit issues if you have any questions.

Update 01/11/2019

We've made the benchmark of Oxford (or Paris) simply. You only need to prepare the oxford dataset images (download from the official website) and run `', see "Note on Oxford5K or Paris6K" for details. Using the default settings in the script, you can easily achieve 0.915 mAP on Oxford5K.


The code base has been tested under TensorFlow 1.5 (CUDA 8.0) to TensorFlow 1.7 (CUDA 9.0), using Python 2.7.12.

First Try

Mirror is extremely easy to use, first download the model files using the following command, suppose you are in the root directory of the project $MIRROR_ROOT

sh ./data/model/

Suppose you have a folder containing the images of the same scene, take fountain-P11 as an example (or any other image folder by your hand), generate the image list by cd into that folder and output absolute image paths to a file:

cd /path/to/fountain-p11
ls -d $PWD/*.png > image_list

Then generate the .npy features:

python retrieval/ --img_list /home/tianwei/Data/fountain-p11/image_list --ckpt_step 720000 --img_size 896 --net resnet-50 --pool MAX --rmac_step 1,3,5 --ckpt_path ./data/model/model.ckpt

Then output all the absolute paths of feature files (*.npy) to a file and query:

ls -d $PWD/*.npy > feature_list
python retrieval/ --feature_list /home/tianwei/Data/fountain-p11/feature_list --out_dir ./output --top 20 --out_dim 256

You should now be able to see a match_pairs file in the output folder, recording the ranking and corresponding similarity scores (in range [0, 1] where 1 means identical) between pairs of images (represented by index).

0 0 1.0
0 1 0.9598151070731027
0 2 0.9315839494977679
0 3 0.9064583369663783
0 4 0.8789794376918247
0 9 0.8777283804757254
0 8 0.868733160836356
0 5 0.8614536830357142
0 10 0.8602364676339286
0 7 0.8555836813790457
0 6 0.8524237496512277
1 1 1.0
1 2 0.9604712350027902
1 0 0.958723885672433
1 3 0.9399954659598214
1 4 0.9106133052280971
... ...

This is the basic usage of this repo. If you cannot get it done, please submit a issue in the issue tracker and let us help you! If you find this work useful in your production environment or research paper, 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},

The following describe our dataset and results presented in the paper.

GL3D dataset

GL3D overview

The GL3D contains 90,590 high-resolution images in 378 different scenes. Each scene contains 50 to 1,000 images with large geometric overlaps, covering urban, rural area, or scenic spots captured by drones from multiple scales and perspectives. Several small objects are also included to enrich the data diversity. Each scene data is reconstructed to generate a triangular mesh model by the state-of-the-art 3D reconstruction pipeline.

Currently we release the 9368 test images for evaluation (see #Test). The full dataset can be found here.


Training code will be released soon.


Please refer to for using the image retrieval pipeline. We release two trained models to demonstrate the use. The googlenet model can be used to reproduce the results in the paper, which achieves 0.758 mAP@200 on GL3D, 0.768 mAP on Oxford5K and 0.820 on Paris6K using the default settings in

We have additionally trained a ResNet-50 model not documented in the original paper, which achieves better performance than GoogleNet on object retrieval datasets (Oxford5K and Paris6K).

Model GL3D (mAP@200) Oxford5K Paris6K Holidays
GoogleNet 0.758 0.768 0.820 0.861
ResNet-50 0.745 0.833 0.805 0.892
ResNet-50 + QE - 0.894 0.858 0.881

Note on GL3D

Download the GL3D test images of size 896 x 896 (around 3.7GB) and unzip them to your GL3D image folder (suppose it is $MIRROR_ROOT/data/gl3d/), and then run the commands for testing. The detailed instruction is given below.

# Download evaluation dataset and unzip
mv ./data/gl3d/
cd data/gl3d/
cd -
# Set up testing configs
net_type=resnet-50 #googlenet
step=720000 #400000 
# Extract features
python retrieval/ --img_list $test_image_list --ckpt_step $step --img_size $img_size \
    --net $net_type --pool $pooling --rmac_step $rmac_step --ckpt_path $ckpt_save_dir --gl3d_root $gl3d_root_dir
# PR-MAC + Top-k
python retrieval/ --feature_list $test_feature_list --out_dir $output_dir --top 200 --out_dim $feature_dim --dataset_root $gl3d_root_dir
# Compute statistics
python retrieval/ $test_image_list $output_dir/match_pairs data/gl3d/gl3d_gt_pairs

Note on Oxford5K or Paris6K

To run the model on Oxford5K or Paris6K, there is one additional step. Suppose you are in the root directory of this repo, you need to first compile the C++ program compute_ap for computing the average precision (AP).

cd cpp
g++ -o compute_ap compute_ap.cpp

The commands are similar to GL3D:

dataset=oxford  #paris 

# extract query features
rm -rf $query_output_path
python retrieval/ --img_list $oxford_image_list --oxford_gt $ground_truth_folder --net $net_type --pool $pooling \
    --ckpt_step $step --ckpt_path $ckpt_save_dir --img_size $img_size --rmac_step $rmac_step --multi_scale 
ls -d $PWD/$query_output_path/*.npy > $query_output_path/query_list

# extract db features
rm -rf $db_output_path
python retrieval/ --output_folder $db_output_path --img_list $oxford_image_list --net $net_type --pool $pooling \
    --ckpt_step $step --ckpt_path $ckpt_save_dir --img_size $img_size --rmac_step $rmac_step --multi_scale 
ls -d $PWD/$db_output_path/*.npy > $db_output_path/feature_list

# query
python retrieval/ --out_dir $output_dir --feature_list $oxford_feature_list \
    --query_list oxford_queries/query_list --top 6500 --out_dim $feature_dim --qe 10 --et 2 #--pca_thresh 0.90 #--rmac --pca_file $pca_file 

# evaluate
python retrieval/ data/$dataset/query_list.txt data/$dataset/image_list.txt $output_dir/match_pairs $ground_truth_folder \
    --output_dir $output_dir

# remove temporary feature files
rm -rf $query_output_path
rm -rf $db_output_path

Note on Holidays

Downloads INRIA Holidays dataset at official website, prepare the required input lists and use the commands in to reproduce the results.

Related Projects

Also checkout the following related geometric learning repositories:

GeoDesc: Learning Local Descriptors by Integrating Geometry Constraints

MVSNet: Depth Inference for Unstructured Multi-view Stereo



You can’t perform that action at this time.