This repository contains the code used in my masters thesis " Resource-Aware Convolutional Nerual Networks for Object Detection using Video Temporal Data" at TUM LIS.

1. c compiler
2. c++11 compiler
3. OpenCV3
4. OpenMP	'Most compilers are shipped with it'
5. cmake
6. Darknet configuartion and weight files

Darknet weight files can be downloaded from here. Differenct weights can be found on Darknet website

Linux: To setup the envirnment for linux user with apt-get based package system follow these steps:

sudo apt-get install build-essentials cmake opencv3 

OSX: To setip the enviornemtn for OSX (macOS) users follow these steps:

/usr/bin/ruby -e "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install)"
brew update
brew install clang-omp
brew install cmake
brew install opencv3
brew ln opencv3 --force

To compile, copy the repo to your machine and using the terminal navgiagte to repo directory

mkdir build && cd build
cmake .. #Linux Users
CC=clang-omp CXX=clang-omp++ cmake .. #OSX Users
make

RAHE support 4 modes of operation:

1. Normal: This a passthrough to YOLO's original detector and processes and single image. Usage:

./RAHE yolo [cfg] [weight] [image-file]

2. YOLO-Stripe: Benchmarks YOLO-Stripe method over the VOC2012 validation set. Describtion in Benchmarking

3. YOLO-Occur: Ouputs the occurence of the number of processed stipes using Stripe-YOLO over VOC2012 validation set. Describtion in Benchmarking

4. Demo: A proof-of-concept real-time object detector. Usage:

./RAHE demo [cfg] [weights]
press S to toggle Sripped mode.

1. Download the VOC2012 validation dataset and the development kit code from here and here, then extract the archives.

2. Copy the val.txt from VOC2012 directory to RAHE root directory. The following commands creates the necessary filenames to be used by Darknet:

    cd ${RAHE_ROOT}
    cp ${VOCdevkit}/VOC2012/ImageSets/Main/val.txt .
    sed 's?^?'`pwd`'${VOCdevkit}/VOC2012/JPEGImages/?; s?$?.jpg?' val > voc.2012.val
   

3. Place voc.2012.val inside build/data/ directory.

4. Benchmarck are designed to be run on multipe PCs to fasten the compuatations.

   a. Strip-YOLO: To becnhmark the accuracy of Stripe-YOLO run the following command:

    ./RAHE stripped [cfg] [weights] [output-base-file] [offset] [range]
   

   b. Occurance: To calculate the occurance rate of the number of stripes in Stripe-YOLO, run the following commands:

    ./RAHE occur [cfg] [weights] [output-base-file] [offset] [range]
   

output-base-file is the output directory + the base file name for the results e.g. results/comp4_det_val_stripped

offset is the number of images to be skipped from VOC2012 set.

range is the number of images to be processed by this PC.

The resulting txt files are to be processed by the scripts avialble in scripts directory. Then to produce mAP and the recall-precession curve, use the MATLAB scripts available in ${VOCdevkit}/VOCcode/ specifically the VOCevaldet.m script.

This code uses three main external libraries:

1. Darknet
2. CppMT
3. RectangleBinPack