camera-relocalisation

camera-relocalisation copied to clipboard

New!!!

We have released a PyTorch implementation of the method for relative camera pose estimation. The code and pre-trained models are available at https://github.com/AaltoVision/RelPoseNet

Source code (Torch, MATLAB), models, and dataset for Camera Relocalization by Computing Pairwise Relative Poses Using Convolutional Neural Network link.

Getting Started

In order to reproduce results presented in Table 1 of the original paper, please do the following:

• Download the 7-Scenes dataset from here.
• Resize all the images such the smaller dimension is 256 and keep the aspect ratio:

find . -name "*.color.png" | xargs -I {} convert {} -resize "256^>" {}

• Download trained weights from here (md5sum: 15d0222e9737c3f558fad2e4d63f48d2).
• From cnn_part folder run:

th main.lua -weights <path/to/downloaded_weights/model_snapshot_7scenes.t7> -dataset_src_path </path/to/7Scenes> -do_evaluation

By default, calculated features would be saved to cnn_part/results/7scenes_res.bin

• To measure localisation performance run matlab filter_pose.m

University Dataset

• The University dataset (~29Gb) is available here (md5sum: 6f512e6c55006c3f6fa0bf3f75f93284).
• Resize images according to 7-Scenes dataset (keeping aspect ratio).
• Download trained weights from here (md5sum: 227caa217653b48ffdf27a9826b838e5).
• From cnn_part folder run:

th main.lua -dataset_name University -weights <path/to/downloaded_weights/model_snapshot_university.t7> -dataset_src_path </path/to/University> -results_filename ./results/university_res.bin -do_evaluation

• To measure localisation performance open filter_pose.m and change pred_file_id to a binary file with estimates, i.e cnn_part/results/university_res.bin

How to cite

If you use this software in your own research, please cite our publication:

@InProceedings{LMKK2017ICCVW,
    author = {Laskar, Zakaria and Melekhov, Iaroslav and Kalia, Surya and Kannala, Juho},
    title = {Camera Relocalization by Computing Pairwise Relative Poses Using Convolutional Neural Network},
    booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
    month = {Oct},
    year = {2017}
}