Steven Puttemans StevenPuttemans

Overview of frequently used docker command

Running dockers

Run your docker with a specific executable and then shut down: sudo nvidia-docker run --rm nvidia/cuda nvidia-smi

Keep running your docker with bash access: sudo nvidia-docker run -ti --rm nvidia/cuda /bin/bash

Building your own docker file

Using FFMPEG a video can be generated from frames using

ffmpeg -r 1/5 -i img%03d.png -c:v libx264 -vf fps=25 -pix_fmt yuv420p out.mp4

Parameters

r = ratio if you want to downscale the video
i = format of the images, adapt it to your needs!

Things to investigate considering the darknet framework

Possibility of actually freezing the layers. This can be done by editing the last layer in the *.cfg file and adding the following parameter l.stopbackward = option_find_int_quiet(options, "stopbackward", 0);.

Improving YOLO for close together objects

Increase resolution of network: https://github.com/AlexeyAB/darknet/blob/b585124e55fcd18a7511d3b2fdc6bc3db5368ef4/cfg/yolo-voc.2.0.cfg#L4
And increase nms: https://github.com/AlexeyAB/darknet/blob/b585124e55fcd18a7511d3b2fdc6bc3db5368ef4/src/detector.c#L459
Also if your images and objects on images both are big, then you will should add additional convolutional layers: https://github.com/AlexeyAB/darknet/blob/b585124e55fcd18a7511d3b2fdc6bc3db5368ef4/src/box.c#L279

This can be done with the following commands

git pull --squash origin_PJREDDIE master
git add --all
git commit -m "Merge upstream PJReddie fixes"
git push -f https://gitlab.com/EAVISE/darknet.git master

Of course, if during the pull merge conflicts appear, you need to solve those first.

	# Rule based filtering of skin pixels
	# (RED>95) && (GREEN>40) && (BLUE>20) && ((max(RED,max(GREEN,BLUE)) - min(RED,min(GREEN,BLUE)))>15) && (abs(RED-GREEN)>15) && (RED>GREEN) && (RED>BLUE);

	result = np.zeros(img_color.shape)

	# Method 1 - naive with loops
	# Remember BGR color space in OpenCV
	for row in range(img_color.shape[0]):
	for col in range(img_color.shape[1]):
	if (img_color[row, col, 2] > 95): # Condition on RED channel

	+++
	# Travel locations
	widget = "custom"
	active = true
	date = "2018-03-23T00:00:00"

	+++

	<head>
	<meta content="text/html; charset=utf-8" http-equiv="Content-Type">

	Saving weights to backup_orchid/(.*)weights\n

	--> (.*) is for changing text
	--> \n to include newline character

	For this enable
	1. use regular expressions
	2. use multi-line matching

	#include <iostream>
	#include "opencv2/opencv.hpp"

	#define DEBUG_OUTPUT 1

	using namespace std;
	using namespace cv;

	Mat own_threshold(Mat conf_map, float lower, float upper){
	Mat temp1 = conf_map > lower;