OpenCV Zoo and Benchmark

A zoo for models tuned for OpenCV DNN with benchmarks on different platforms.

Guidelines:

Clone this repo to download all models and demo scripts:

# Install git-lfs from https://git-lfs.github.com/
git clone https://github.com/opencv/opencv_zoo && cd opencv_zoo
git lfs install
git lfs pull

Models & Benchmark Results

Model	Task	Input Size	INTEL-CPU (ms)	RPI-CPU (ms)	JETSON-GPU (ms)	KV3-NPU (ms)	D1-CPU (ms)
YuNet	Face Detection	160x120	1.45	6.22	12.18	4.04	86.69
SFace	Face Recognition	112x112	8.65	99.20	24.88	46.25	---
LPD-YuNet	License Plate Detection	320x240	---	168.03	56.12	154.20*
DB-IC15	Text Detection	640x480	142.91	2835.91	208.41	---	---
DB-TD500	Text Detection	640x480	142.91	2841.71	210.51	---	---
CRNN-EN	Text Recognition	100x32	50.21	234.32	196.15	125.30	---
CRNN-CN	Text Recognition	100x32	73.52	322.16	239.76	166.79	---
PP-ResNet	Image Classification	224x224	56.05	602.58	98.64	75.45	---
MobileNet-V1	Image Classification	224x224	9.04	92.25	33.18	145.66*	---
MobileNet-V2	Image Classification	224x224	8.86	74.03	31.92	146.31*	---
PP-HumanSeg	Human Segmentation	192x192	19.92	105.32	67.97	74.77	---
WeChatQRCode	QR Code Detection and Parsing	100x100	7.04	37.68	---	---	---
DaSiamRPN	Object Tracking	1280x720	36.15	705.48	76.82	---	---
YoutuReID	Person Re-Identification	128x256	35.81	521.98	90.07	44.61	---
MP-PalmDet	Palm Detection	256x256	15.57	168.37	50.64	145.56*	---
MP-HandPose	Hand Pose Estimation	256x256	20.16	148.24	156.30	663.77*	---

*: Models are quantized in per-channel mode, which run slower than per-tensor quantized models on NPU.

Hardware Setup:

INTEL-CPU: Intel Core i7-5930K @ 3.50GHz, 6 cores, 12 threads.
RPI-CPU: Raspberry Pi 4B, Broadcom BCM2711, Quad core Cortex-A72 (ARM v8) 64-bit SoC @ 1.5GHz.
JETSON-GPU: NVIDIA Jetson Nano B01, 128-core NVIDIA Maxwell GPU.
KV3-NPU: Khadas VIM3, 5TOPS Performance. Benchmarks are done using quantized models. You will need to compile OpenCV with TIM-VX following this guide to run benchmarks. The test results use the per-tensor quantization model by default.
D1-CPU: Allwinner D1, Xuantie C906 CPU (RISC-V, RVV 0.7.1) @ 1.0GHz, 1 core. YuNet is supported for now. Visit here for more details.

Important Notes:

The data under each column of hardware setups on the above table represents the elapsed time of an inference (preprocess, forward and postprocess).
The time data is the median of 10 runs after some warmup runs. Different metrics may be applied to some specific models.
Batch size is 1 for all benchmark results.
--- represents the model is not availble to run on the device.
View benchmark/config for more details on benchmarking different models.