CLIDD: Cross-Layer Independent Deformable Description for Efficient and Discriminative Local Feature Representation

We introduce Cross-Layer Independent Deformable Description (CLIDD), a high-performance local feature representation method designed for scalability. It provides a diverse range of models, from an ultra-compact 4,252 (0.004M) parameter variant to high-performance configurations exceeding 200 FPS on edge devices, delivering a robust and efficient solution for image description and matching.

The visualization below shows the A48 configuration tested on the KITTI dataset.

Model Zoo

Available model variants are listed below. FPS results are measured on an NVIDIA Jetson Orin-NX.

Model	Dim	MP	FPS
A48	48	0.004	881.1
N64	64	0.019	842.7
T64	64	0.043	803.5
S64	64	0.100	746.4
M64	64	0.168	625.1
L64	64	0.347	591.4
G128	128	2.254	475.4
E128	128	3.508	431.3
U128	128	4.400	281.4

Install Dependencies

Set up the environment by installing the required packages:

pip install -r requirements.txt

Usage

Run the demo by specifying the input source and model:

python demo_seq.py [camera / PATH_TO_VIDEO_FILE] [Model]

Example for camera input with the A48 model:

python demo_seq.py camera A48

Note: Initial execution or significant changes in the number of points may cause brief stutters as Triton re-optimizes the kernels.