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 |
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
- Install CUDA, CUDNN
- Install OpenCV
- Install QT by terminal
- Make sure to install OpenCV, CUDA, CUDNN perfectly and try to build DarkNet
- Download the weight files link.
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