gplast/dronet

DroNet: Efficient convolutional neural network detector for Real-Time UAV applications

DroNet: Efficient Convolutional Neural Network Detector for Real-Time UAV Applications

Implementation of the CNN Car Detector proposed on DroNet paper using Darknet Framework.

Dependencies

OpenCV

CUDA(Optional)

Build

In order to build DarkNet library and use our DroNet CNN you need to build OpenCV using these Steps.

nano Makefile # run this with sudo if you have permission error

To build DroNet on Mac OS please comment out line 22 and uncomment line 23 and update the path of GCC

• GPU=0 - enable/disable GPU (To use GPU modify line 50 and 52 with include and lib64 CUDA path)
• CUDNN=0 - enable/disable CUDNN
• OPENCV=1 - enable/disable OpenCV
• OPENMP=1 - enable/disable Multi-Processing on CPU
• DEBUG=0 - enable/disable Debug mode (Never used)

make -j

DroNetV1 - Works better with low altitudes

./darknet detector demo car.data cfg/DroNet_car.cfg results/DroNet_car.weights Car_Parking.mov -thresh 0.4
./darknet detector demo car.data cfg/DroNet_car.cfg results/DroNet_car.weights Car_Road.MOV -thresh 0.4

DroNetV3 - Works better with higher altitudes (Recommended .cfg input 1024 x 1024)

./darknet detector demo car_ped.data cfg/DroNetV3_car.cfg results/DroNetV3_car.weights Car_Crossroad.mp4

dronet's Issues

training data

HI, can you share your training data? thanks

training weights 250mb

hi @gplast , I'm replicating your repo for training a yolo model on custom data. My target is to get the weights less than 1MB in size. But after the training is finished I see they are around 250 MB. Did I miss anything here? or should I do another step? Can you please help me?

training procedure

Hi ,
First of all thanks for this great work !
I would like to replicate your system for a custom object detector.
What were the steps you took to train.
I assume that batch size was larger than 1. Did you pretrain your network in anyway or did you train directly on your data? Did you optimize the anchors in anyway ? Anything else that was significant in stabilizing the learning or improving accuracy?

I am also curious to hear if you tried a version with only 4 down sampling steps and 4x fewer features after that to deal with smaller objects? or using a second "yolo" layer like in tiny yolo v3.

Thanks,
Dan