Skip to content

A real-time object detection app based on lightDenseYOLO Our lightDenseYOLO is the combination of two components: lightDenseNet as the CNN feature extractor and YOLO v2 as the detection module

Notifications You must be signed in to change notification settings

LongChenCV/lightDenseYOLO

 
 

Repository files navigation

lightDenseYOLO

A real-time object detection app based on lightDenseYOLO Our lightDenseYOLO is the combination of two components: lightDenseNet as the CNN feature extractor and YOLO v2 as the detection module

lightDenseYOLO was trained with two famous object detection datasets (MS COCO and Pascal VOC 07+12)

CNN architecture Training Data mAP Processing time
lightDenseYOLO (2 blocks) VOC 70.7 20 ms ~ 50 fps
lightDenseYOLO (4 blocks) VOC 77.1 28 ms ~ 35.8 fps
YOLO v2 VOC 75.4 30 ms ~ 33 fps
Faster-RCNN + Resnet 101 VOC 78.9 200 ms ~ 5 fps
MobileNets + SSD VOC 73.9 80 ms ~ 12.5 fps
lightDenseYOLO (2 blocks) VOC + COCO 79.0 20 ms ~ 50 fps
lightDenseYOLO (4 blocks) VOC + COCO 82.5 28 ms ~ 35.8 fps
YOLO v2 VOC + COCO 81.5 30 ms ~ 33 fps
Faster-RCNN + Resnet 101 VOC + COCO 83.8 200 ms ~ 5 fps
MobileNets + SSD VOC + COCO 76.6 80 ms ~ 12.5 fps

Usage

Requirements:

  • Ubuntu 16.04+
  • C++ 11 complier
  • QT 5.7.0 and QT Creator 4.0.2
  • Open CV 3.2 +
  • GPU (NVIDIA Gefore 1070+)

Installation

Build:

  • Create a new Qt project,
  • Copy all existing file in the project folder.
  • Create a new folder named data next to the created project folder and copy the weight file and the cfg file there.
  • Hit build and run.

Example video

References

1.LightDenseYOLO: A Fast and Accurate Marker Tracker for Autonomous UAV Landing by Visible Light Camera Sensor on Drone

2.YOLO 9000

3.Faster-RCNN

4.MobileNets

5.MobileNets+SSD

About

A real-time object detection app based on lightDenseYOLO Our lightDenseYOLO is the combination of two components: lightDenseNet as the CNN feature extractor and YOLO v2 as the detection module

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C++ 98.2%
  • QMake 1.8%