ALIKED: A Lighter Keypoint and Descriptor Extraction Network via Deformable Transformation

ALIKED is an improvement on ALIKE, which introduces the Sparse Deformable Descriptor Head (SDDH) to efficiently extract deformable descriptors. Compared with ALIKE, ALIKED can extract more robust descriptors in a more efficient way. The technical details are described in this paper.

Xiaoming Zhao, Xingming Wu, Weihai Chen, Peter C.Y. Chen, Qingsong Xu, and Zhengguo Li, "ALIKED: A Lighter Keypoint and Descriptor Extraction Network via Deformable Transformation", IEEE Transactions on Instrumentation & Measurement, 2023.

If you use ALIKED in academic work, please cite:

@article{Zhao2023ALIKED,
    title = {ALIKED: A Lighter Keypoint and Descriptor Extraction Network via Deformable Transformation},
    url = {https://arxiv.org/pdf/2304.03608.pdf},
    doi = {10.1109/TIM.2023.3271000},
    journal = {IEEE Transactions on Instrumentation & Measurement},
    author = {Zhao, Xiaoming and Wu, Xingming and Chen, Weihai and Chen, Peter C. Y. and Xu, Qingsong and Li, Zhengguo},
    year = {2023},
    volume = {72},
    pages = {1-16},
}

@article{Zhao2022ALIKE,
    title = {ALIKE: Accurate and Lightweight Keypoint Detection and Descriptor Extraction},
    url = {http://arxiv.org/abs/2112.02906},
    doi = {10.1109/TMM.2022.3155927},
    journal = {IEEE Transactions on Multimedia},
    author = {Zhao, Xiaoming and Wu, Xingming and Miao, Jinyu and Chen, Weihai and Chen, Peter C. Y. and Li, Zhengguo},
    month = march,
    year = {2022},
}

1. Prerequisites

The required packages are listed in the requirements.txt :

pip install -r requirements.txt

Build custom_ops:

cd custom_ops
sh build.sh

2. Pretrained models

The pretrained ALIKED models are provided in models/ .

3. Demo

a) image pair demo

Example:

python demo_pair.py assets/st_pauls_cathedral

python demo_pair.py assets/piazza_san_marco

Usage:

$ python demo_pair.py -h
usage: demo_pair.py [-h] [--model {aliked-t16,aliked-n16,aliked-n16rot,aliked-n32}] 
                    [--device DEVICE] [--top_k TOP_K] [--scores_th SCORES_TH] 
                    [--n_limit N_LIMIT] input

ALIKED image pair Demo.

positional arguments:
  input                 Image directory.

options:
  -h, --help            show this help message and exit
  --model {aliked-t16,aliked-n16,aliked-n16rot,aliked-n32}
                        The model configuration
  --device DEVICE       Running device (default: cuda).
  --top_k TOP_K         Detect top K keypoints. -1 for threshold based mode, >0 for top K mode. (default: -1)
  --scores_th SCORES_TH
                        Detector score threshold (default: 0.2).
  --n_limit N_LIMIT     Maximum number of keypoints to be detected (default: 5000).

b) sequence demo

Example:

python demo_seq.py assets/tum

Usage:

$ python demo_seq.py -h
usage: demo_seq.py [-h] [--model {aliked-t16,aliked-n16,aliked-n16rot,aliked-n32}] 
                   [--device DEVICE] [--top_k TOP_K] [--scores_th SCORES_TH] 
                   [--n_limit N_LIMIT] [--no_display] input

ALIKED sequence Demo.

positional arguments:
  input                 Image directory or movie file or "camera0" (for webcam0).

options:
  -h, --help            show this help message and exit
  --model {aliked-t16,aliked-n16,aliked-n16rot,aliked-n32}
                        The model configuration
  --device DEVICE       Running device (default: cuda).
  --top_k TOP_K         Detect top K keypoints. -1 for threshold based mode, >0 for top K mode. (default: -1)
  --scores_th SCORES_TH
                        Detector score threshold (default: 0.2).
  --n_limit N_LIMIT     Maximum number of keypoints to be detected (default: 5000).
  --no_display          Do not display images to screen. Useful if running remotely (default: False).

4. Efficiency and performance

Image matching & multiview reconstruction

Relocalization

5. Limitations

The aliked-n16rot is trained with rotation augmentation, it can handle some image rotation and is more stable on viewpoint change. But it can still have difficulty on large image rotations.
ALIKED is detector-based, and it is designed as light-weight as possible, so it is not comparable with the detector-free methods.
The training code: I am sorry that I am not allowed to publish the training code of ALIKED, you can refer to the training code of ALIKE.

For more details, please refer to the paper.

Shiaoming/ALIKED