Benchmarking on Edge Devices

Environment

General points

All benchmark are performed one-by-one with restarting kernel before running each cell.

Raspberry Pi 5

1. Raspberry Pi OS Lite (64-bit).

2. Python 3.11.

3. Original source code and packages (poetry).

4. Active cooling.

Orange Pi 5B

1. Ubuntu 22.04.

2. Python 3.11.

3. Original source code and packages (poetry).

4. Active cooling.

Jetson Nano 4Gb

1. Ubuntu 20.04 Q-engineering.

2. Python 3.8.

3. Preinstalled Pytoch 1.13 with CUDA support.

4. One-time setup with the manually installed packages.

5. Slightly modified source code to fit Python 3.8 syntax and Pytorch 1.13, stored in a dedicated folder:

   • scaled_dot_product_attention is implemented in place;
   • minor syntax changes.

SuperPoint

descriptor_dim = 256, num_keypoints = 512, num_iterations = 10.

256x256

Model	Raspberry Pi 5
SuperPoint	4954
SuperPointONNX	296
superpoint.onnx	147
superpoint_I256_D256_K512.onnx	153

1024x1024

Model	Raspberry Pi 5
SuperPoint	5107
SuperPointONNX	4969
superpoint.onnx	2860

LightGlue

descriptor_dim = 256, num_keypoints = 512, num_iterations = 10.

Model	Raspberry Pi 5
LightGlueONNX	766
superpoint_lightglue_fullgraph.onnx	516
superpoint_lightglue_fullgraph_I256_D256_K512.onnx.onnx	494

Detectors

Benchmark is performed on real images, num_iterations = 10.

Model	Raspberry Pi 5
LightGlueDetector	11223
LightGlueORTDetector	15144
LightGlueORTDetector, 512 keypoints	7017