Home

Awesome

2022.7.14:Optimize loss, adopt IOU aware based on smooth L1, and the AP is significantly increased by 0.7

:zap:FastestDet:zap:

DOI image image image

Evaluating indicator/Benchmark

NetworkmAPval 0.5mAPval 0.5:0.95ResolutionRun Time(4xCore)Run Time(1xCore)Params(M)
yolov5s56.8%37.4%640X640395.31ms1139.16ms7.2M
yolov6n-30.8%416X416109.24ms445.44ms4.3M
yolox-nano-25.8%416X41676.31ms191.16ms0.91M
nanodet_m-20.6%320X32049.24ms160.35ms0.95M
yolo-fastestv1.124.40%-320X32026.60ms75.74ms0.35M
yolo-fastestv224.10%-352X35223.8ms68.9ms0.25M
FastestDet25.3%13.0%352X35223.51ms70.62ms0.24M

Improvement

Multi-platform benchmark

EquipmentComputing backendSystemFrameworkRun time(Single core)Run time(Multi core)
Radxa rock3aRK3568(arm-cpu)Linux(aarch64)ncnn70.62ms23.51ms
Radxa rock3aRK3568(NPU)Linux(aarch64)rknn28ms-
QualcommSnapdragon 835(arm-cpu)Android(aarch64)ncnn32.34ms16.24ms
Inteli7-8700(X86-cpu)Linux(amd64)ncnn4.51ms4.33ms

How to use

Dependent installation

Test

<div align=center> <img src="https://github.com/dog-qiuqiu/FastestDet/blob/main/result.png"> /> </div>

How to train

Building data sets(The dataset is constructed in the same way as darknet yolo)

Build the training .yaml configuration file

Train

Evaluation

Deploy

Export onnx

Export torchscript

NCNN

onnx-runtime

Citation

Reference