<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Makerere AI Lab - Medium]]></title>
        <description><![CDATA[The AI &amp; Data Science research group at Makerere University specialises in the application of artificial intelligence and data science - including, for example, methods from machine learning, computer vision and predictive analytics - to problems in the developing world. - Medium]]></description>
        <link>https://medium.com/ai-research-lab-kampala?source=rss----ca834513b05---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Makerere AI Lab - Medium</title>
            <link>https://medium.com/ai-research-lab-kampala?source=rss----ca834513b05---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 15:02:34 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/ai-research-lab-kampala" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Getting an object detector/tracker up and running on Raspberry Pi 3 (Part II)]]></title>
            <link>https://medium.com/ai-research-lab-kampala/getting-an-object-detector-tracker-up-and-running-on-raspberry-pi-3-part-ii-288421050a60?source=rss----ca834513b05---4</link>
            <guid isPermaLink="false">https://medium.com/p/288421050a60</guid>
            <category><![CDATA[rasberry-pi]]></category>
            <dc:creator><![CDATA[Flavia Ninsiima]]></dc:creator>
            <pubDate>Thu, 13 Sep 2018 16:48:04 GMT</pubDate>
            <atom:updated>2018-09-13T16:51:10.933Z</atom:updated>
            <content:encoded><![CDATA[<p>Hope you enjoyed <a href="https://medium.com/ai-research-lab-kampala/getting-an-object-detector-tracker-up-and-running-on-raspberry-pi-3-part-i-d246efe377d6">Part I</a> of this series. Let’s move on to the <a href="https://www.raspberrypi.org/documentation/configuration/raspi-config.md">Pi setup</a>. Recall that this setup is for the <strong>Raspberry Pi 3</strong> model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Bb-FCgaBrVDCJORCu78PtA.jpeg" /></figure><p>Things we will configure:</p><ul><li>Storage expansion to utilize the sdcard</li><li>PiCamera</li><li>Swap memory (optional)</li></ul><ol><li><strong>Expanding FileSystem</strong></li></ol><p>Why is it necessary? Because it’s so easy to run out of memory! The official answer is — <em>to ensure that all of the SD card storage is available to the OS</em>. Ben learnt this the hard way after he tried to build opencv from source and run out of memory. So let’s expand the FileSystem storage to enable it to access the sdcard. In the terminal enter</p><pre>sudo raspi-config.</pre><p>Choose Option 5: <em>Advanced Options</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9XLlAa9bQdgvFjqGpvWldQ.jpeg" /></figure><p>The menu, which appears has as first menu item: <em>Expand Filesystem ensures that all of the SD card storage is available to the OS</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5KfYwI_79cGkLWFY-2duiA.jpeg" /></figure><p>REBOOT!</p><p><strong>2. Attach PiCamera and Configure it</strong></p><p>Setting up the camera board is not so hard. The guy in this <a href="https://youtu.be/GImeVqHQzsE">video</a> does a very simple job of explaining how to insert it into the Pi. Insert it when the Pi is<em> </em><strong><em>off</em></strong><em>.</em></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FGImeVqHQzsE%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DGImeVqHQzsE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FGImeVqHQzsE%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/2ee55b7bd4384054850b842d776d5003/href">https://medium.com/media/2ee55b7bd4384054850b842d776d5003/href</a></iframe><p>Let’s also configure PiCamera as directed in the <a href="https://projects.raspberrypi.org/en/projects/getting-started-with-picamera/4">website</a>. Boot your PI, run the sudo raspi-config command and select Option 3:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZmLQb0yvEM0_p6VpWbr_mw.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*heRTGkA7r8Sqr_xKMW7E7w.jpeg" /><figcaption>Select option P1 to Enable the Camera</figcaption></figure><p>Just to show you what this looks like on my end!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vKxx_jRjVTJ68U6D_Dvu0Q.jpeg" /></figure><p>So my camera is facing the door of our lab and it will capture video frames and the code i will refer to will use background subtraction to detect presence of a new object in the frame.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6NGC0Oeq0kkgi8Fghghs6Q.jpeg" /></figure><p><strong>Note:</strong> This time round we decided to connect our Pi to a projector using an HDMI cable. In the first part, we used a <strong>VGA to HDMI</strong> cable to connect to a Monitor for display. We also demonstrated how you can <strong>ssh</strong> into the Pi and access its desktop and terminal using something called <strong>VNC Connect</strong>. Both these options are viable and can be used for this purpose. You could even use your T.V!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t4nDKJCMCks57M6DXDKDMw.jpeg" /><figcaption>HDMI Cable</figcaption></figure><p>Let’s move on to the motion detection</p><p><strong>Disclaimer</strong>: Though the Pi setup is entirely ours, the code that does the detection is borrowed from <a href="https://www.pyimagesearch.com/2015/06/01/home-surveillance-and-motion-detection-with-the-raspberry-pi-python-and-opencv/">this pyimage tutorial</a> by Adrian Rosebrock.</p><ol><li>Libraries that you will need:</li></ol><ul><li>Python</li><li>Opencv</li><li>Imutils</li></ul><p>2. The code you will need</p><p><a href="https://github.com/fninsiima/motion-detection-with-raspberrypi3">https://github.com/fninsiima/motion-detection-with-raspberrypi3</a></p><p>To run the code:</p><pre>$ python pi_surveillance.py --conf conf.json</pre><p>Let’s see what this looks like when we run it. On start, <em>Room Status: Unoccupied</em>! Over subsequent frames, some background subtraction is done and when a substantial difference is noted, the result is <em>Room Status: Occupied</em>! Adrians’ code performs some contour detection. Wherever the green boxes are is where some new motion has been detected! Case in point the third frame where <a href="https://medium.com/u/b18ce0e9e9b0">solomon nsumba</a> turns his head!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ba1DBBsXROynLQxWut8_LQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3i0yVVNWznFbO3GB8KhqaA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TBCPGYXsPuf1E_eZu2HV3w.jpeg" /><figcaption>Detecting motion</figcaption></figure><p>To see a video of how this works:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F9UxP3ge_5vM%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D9UxP3ge_5vM&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F9UxP3ge_5vM%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/7373e2ae6163daacc1e9aba275c084cc/href">https://medium.com/media/7373e2ae6163daacc1e9aba275c084cc/href</a></iframe><p>This is the end of the post. Thank you for joining for this series.</p><p><strong>Important links for further reading</strong></p><ul><li><a href="https://www.pyimagesearch.com/2015/06/01/home-surveillance-and-motion-detection-with-the-raspberry-pi-python-and-opencv/">Home surveillance and motion detection with the Raspberry Pi, Python, OpenCV, and Dropbox - PyImageSearch</a></li><li><a href="https://wpitchoune.net/tricks/raspberry_pi3_increase_swap_size.html">Raspberry PI - increase swap size</a></li><li><a href="https://www.raspberrypi.org/documentation/configuration/raspi-config.md">raspi-config</a></li><li><a href="https://projects.raspberrypi.org/en/projects/getting-started-with-picamera/5">Getting started with picamera - Introduction | Raspberry Pi Projects</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=288421050a60" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-research-lab-kampala/getting-an-object-detector-tracker-up-and-running-on-raspberry-pi-3-part-ii-288421050a60">Getting an object detector/tracker up and running on Raspberry Pi 3 (Part II)</a> was originally published in <a href="https://medium.com/ai-research-lab-kampala">Makerere AI Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Thursday Nuggets: Intellectual Property Laws and Artificial Intelligence.]]></title>
            <link>https://medium.com/ai-research-lab-kampala/thursday-nuggets-intellectual-property-laws-and-artificial-intelligence-9de6216c2079?source=rss----ca834513b05---4</link>
            <guid isPermaLink="false">https://medium.com/p/9de6216c2079</guid>
            <category><![CDATA[intellectual-property]]></category>
            <category><![CDATA[law]]></category>
            <category><![CDATA[law-firm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[computer-science]]></category>
            <dc:creator><![CDATA[AIR Lab]]></dc:creator>
            <pubDate>Thu, 06 Sep 2018 15:12:52 GMT</pubDate>
            <atom:updated>2018-09-06T15:20:51.782Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong><em>Speakers: Daniel Bill Opio, Baguma Moses, Emmanuel Okiror </em></strong>from<strong><em> CyberLine Law Initiative.</em></strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/945/1*CMhxl88pmZEb5jfvbhPlLw.jpeg" /></figure><p><strong>Disclaimer:</strong> The Thursday Nuggets are a series of blog posts that summarize the seminars held every Thursday at the AI Research Lab at Makerere University. The goal of this series is to provide a platform for future reference in regards to the content discussed during these seminars.</p><p>Artificial Intelligence is pushing innovation in new ways and accelerating with technological advances in computing power, data and algorithms. Many companies are now using machine learning methods ranging from startups to long-established institutions. Intellectual property (IP) protects creativity and encourages innovation. As such, it is important for the companies, entrepreneurs and investors to be aware of the Key IP considerations as applied to innovation and AI. We had the privilege of hosting speakers from the <a href="http://www.cyberline.org/">CyberLine Law Initiative</a> who talked to us about the IP laws in Uganda and how best we can harness them.</p><p><strong>What is Intellectual Property?</strong></p><p>Intellectual Property (IP) refers to creations of the mind, such as inventions; literary and artistic works; designs; and symbols, names and images used in commerce. Intellectual Property rights grant the owner of the work exclusive rights to exploit and benefit from his/her creation.</p><p><strong>Why protect Intellectual Property?</strong></p><p>To provide an incentive to the innovators to be able to benefit from the result of their endeavor.<br>The legal protection of innovations encourages the commitment of additional resources for further innovation.The promotion and protection of intellectual property spurs economic growth, creates jobs and enhances the quality and enjoyment of life.</p><p><strong>Forms of Intellectual Property</strong></p><p>The most common forms of Intellectual Property include: Copyright, trademarks, patents, utility models, industrial designs and geographical indications. IP relates to intangible assets, including inventions, brands, new technologies, source code and artistic works.</p><p><strong>Copyright:</strong> Mr. Baguma explained that copyrights gives protection to an expression of an idea and not just an idea. It protects an idea that has been put to expression. He added that in Uganda, copyright protects the creator for the period of their lifetime plus 50 years after their death.</p><p><strong>Trademarks:</strong> The CyberLine lawyers also hinted on trademarks which are described as “ A sign or mark that distinguishes one’s product from from another. An example of a trademark is the taxify and SafeBoda logos used as each company’s trademark”. A trademark can be in the form of a sign, sound or any thing used in branding.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/945/1*0J4H4KS41C_y-8zMyofXeA.jpeg" /><figcaption>An example of Trademarks are the Uber, SafeBoda and Taxify Logos</figcaption></figure><p><strong>Patents: </strong>Patents provide a time-limited protection for an invention. A patent entitles the patent owner to the exclusive right to make, use and sell his or her invention in exchange for full and clear disclosure on how to work the invention. They provide a mechanism to exclude others from making, using or selling the patented technology, which may help companies obtain or maintain market share, and protect research and development investments. Patents can provide a competitive advantage, and may also be used defensively as a negotiation tool. Patent publications can also be cited against subsequently filed applications to prevent grant.</p><h3><strong>Conclusion</strong></h3><p>We had a great discussion on how IP laws are relevant in the ever growing field of Artificial Intelligence and Machine Learning. This Blog post is just the tip of a bigger iceberg that is IP laws. This field in the global innovation economy definitely demands for intellectual property (IP) titles such as: Copyright, trademarks, patents, industrial design (especially for user interfaces) and utility models though rapidly becoming complex given the nature of AI.</p><p>Nonetheless, the Cyber Law Initiative (CyberLine) legal team led by Daniel <strong>Bill Opio, Moses Baguma and Emmauel Okiror</strong> have taken the initiative under the CyberLine Techubation Hub to conduct a mobile move in legaltech training on matters of the relationship between AI and IP and assist AI developers acquire the requisite requirements for attainment of the rights with emphasis on the importance of acquisition of such rights. The need revolves around protecting trade secrets, exploring exclusive economic rights to the exclusion of others in a bid to spur innovation and economic growth. As a result, CyberLine is on the journey of positioning itself as a Collective Management Organization (CMO) responsible for protecting and pursuing the economic IP rights of Software developers. The future is indeed promising in an atmosphere where CyberLine is not just the light at the end of the AI tunnel but the light in the AI tunnel.</p><p><strong>Prepared By: Benjamin Akera.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9de6216c2079" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-research-lab-kampala/thursday-nuggets-intellectual-property-laws-and-artificial-intelligence-9de6216c2079">Thursday Nuggets: Intellectual Property Laws and Artificial Intelligence.</a> was originally published in <a href="https://medium.com/ai-research-lab-kampala">Makerere AI Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Getting an object detector/tracker up and running on Raspberry Pi 3 (Part I)]]></title>
            <link>https://medium.com/ai-research-lab-kampala/getting-an-object-detector-tracker-up-and-running-on-raspberry-pi-3-part-i-d246efe377d6?source=rss----ca834513b05---4</link>
            <guid isPermaLink="false">https://medium.com/p/d246efe377d6</guid>
            <category><![CDATA[raspberry-pi]]></category>
            <dc:creator><![CDATA[Flavia Ninsiima]]></dc:creator>
            <pubDate>Mon, 03 Sep 2018 08:26:31 GMT</pubDate>
            <atom:updated>2018-09-03T08:26:30.519Z</atom:updated>
            <content:encoded><![CDATA[<p>Recently we were tasked with finding a way to install a person detector at our lab — a tracker that would possibly count who enters and leaves the lab. This was meant to be a fun experiment of sorts with machine learning and IoT. You can see my partner-in-crime <a href="https://medium.com/u/7cf731d10f96">Benjamin Akera</a>(2nd left) in the photo below, as well as the requester (extreme right).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/436/1*pQa0fCDgI2cqgdE0kmdWbw.jpeg" /><figcaption>Detecting persons present in an image. Picture courtesy of a <a href="http://www.datascienceafrica.com/">dsa2018</a> conference attendee.</figcaption></figure><p><strong>Materials</strong>:</p><ul><li>raspberry pi 3 [Model B]</li><li>VGA to HDMI cable</li><li>monitor/desktop/laptop</li><li>keyboard and mouse</li><li>internet connection</li><li>micro SDcard at least 8gb</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hHaLgEz5xfrHN6LNJoFO-A.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Bb-FCgaBrVDCJORCu78PtA.jpeg" /><figcaption>VGA to HDMI Cable(left)</figcaption></figure><p>The first step is to install an operating system on the raspberry Pi, now there are many options you could choose from and the most popular being Raspbian OS. In our case we chose to use Ubuntu MATE (mah-tay)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/822/1*1bFNxe6byi7fVYJ6sjHCXQ.png" /><figcaption>Picture courtesy of <a href="https://snapcraft.io/ubuntu-mate-welcome">Snapcraft</a></figcaption></figure><p><strong>Step 1:</strong> The first thing you have to do is to format your SDcard just to make sure you’re starting on a clean slate. Ensure to backup any files you may have on it. Ideally you can use a sdcard reader but we didn’t have one so we improvised by inserting the card into our phones and connecting to our laptop. Even if the disk drive name doesn’t appear in windows explorer under My Computer, Windows provides a great utility called Disk Management which enables you to format the sdcard. Simply right click on the partition you would like to format and choose <em>Format</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*n-8h_tAKMhlrvQ5zY5Mo2A.jpeg" /><figcaption>Disk management option to format hidden partitions that can’t be viewed in ‘My Computer’</figcaption></figure><p><strong>Step 2:</strong> <a href="https://ubuntu-mate.org/raspberry-pi/ubuntu-mate-16.04.2-desktop-armhf-raspberry-pi.img.xz">Download</a> Ubuntu Mate and flash it to your sdcard. For this test, we downloaded ubuntu-mate-16.04.2-desktop-armhf-raspberry-pi.img and used <a href="https://etcher.io/">Etcher</a> to flash to the sdcard. The flashing and validating process will take a while.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/806/1*Dnj_ZZvAQQIJ3K5iMIJ9rw.jpeg" /></figure><p>For initial access to the Pi, we will employ a VGA to HDMI cable. Also a monitor, keyboard and mouse come in handy!</p><p>Step 1. Insert the HDMI bit of the connector into the Pi and plug the VGA end into the monitor.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x0WEbnwbk_KLopEGGyCYRw.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rqL_TocX_1qLhhfYLIbvtQ.jpeg" /></figure><p>Step 2. Insert the keyboard and mouse to the Pi via USB. Pi’s have usb ports into which you can insert your peripherals.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ln_MtPEHB-AzRZ_jyTmqWA.jpeg" /></figure><p>Step 3. Insert your sdcard into the Pi and power up the Pi. One way to power up the Pi is to insert a USB cable in it and connect to your laptop. However sometimes, the USB ports may not provide sufficient power, especially if you notice a lighting icon on the monitor, so it’s best to manually connect the USB cable from the Pi to a plug used to charge normal smart phones. Let’s not forget the wired connection if you would like to set it up now</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GJm3h7G7wWQnfpbBRLZlsQ.jpeg" /></figure><p>VOILA!!! You should be able to see the start up of Ubuntu MATE. After start up is done, in the terminal you can run $sudo raspi-config to access/enable different options.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S2fVS-2Z2OBYIjMykRGlbA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p1XV7RysczlZ0qrthNmE5w.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3ggLUbUq7BaFA9lw_JX60Q.jpeg" /></figure><p>Next, we need a way to ssh into our raspberry pi. We do this such that we can access the Pi desktop UI and or terminal from any device as long as both the Pi and our device share the same ethernet/wifi connection. After this process is complete, we won’t need the VGA to HDMI cable to connect the Pi to a monitor. But first, leave in the VGA to HDMI cable until we finish setting up ssh access.</p><p>Step 1. Choose whether to connect to the Pi through wifi connection or LAN and set it up. What does this mean? In the following steps when we set up VNC connections for ssh access, our Pi and our laptop/device will need to be connected to the same networks. Therefore if Pi is on a network/internet connection with an APN called “free_wifi”, then the laptop should also be connected to “free_wifi”.</p><p>Step 2. Download VNC Viewer for the operating system of your device (the device you will use to ssh into the Pi), for our case it was Windows.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*g9VRIdoBttPBpfjk_JhKEw.jpeg" /></figure><p>Step 3. Download VNC Connect for the pi and $sudo dpkg -i path/to/vnc_connect to install it. To download VNC, we can use the default firefox browser that comes with Ubuntu MATE. HOWEVER, you have to setup/fix the timezone of the system or else firefox continuously returns errors and web pages will not load.</p><p>Step 4. Fire up VNC Connect on the Pi terminal(Ctrl+Alt+t) using $ vncserver-x11</p><p>VCN Server application pops up and should provide an ip address to the effect of something like 192.168.1.x.</p><p>Step 5. Fire up VNC viewer from your laptop and type in this ip address. Enter the username and password if you set up one using vncserver. If successful, you should now be able to access the desktop environment of your Pi from your device.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/901/1*odFHvFMaYwV6NiKM54GPgA.jpeg" /></figure><p>Next, we must ensure that whenever we boot up our Pi, vcnserver is running and enabled i.e.. “<a href="http://www.google.com/search?q=ubuntu%20launch%20vnc%20server%20on%20startup">ubuntu launch vnc server on startup</a>”.</p><p>Step 1: In Pi terminal run</p><p>systemctl enable vncserver-x11-serviced.service #systemd <br> update-rc.d vncserver-x11-serviced defaults #initd</p><p>However, this is bound to give you errors stating that script init.d doesnt exist. So copy the contents from this file <a href="https://github.com/fninsiima/medium-files/blob/master/vncserver">https://github.com/fninsiima/medium-files/blob/master/vncserver</a> into /etc/init.d/vncserver. The easiest way to do it is to copy it to your clipboard, run sudo -i &amp;&amp; cat &gt; /etc/init.d/vncserver &amp;&amp; exit in a terminal, paste it in, and type CTRL-D. Be sure to change the USER, NAME, GEOMETRY, variable to whatever user you want the VNC server to run under. Make the script executable with sudo chmod +x /etc/init.d/vncserver then, run sudo update-rc.d vncserver defaults. To start the server without rebooting, run sudo /etc/init.d/vncserver start</p><p>Once this is up and running, you can reboot the Pi without the VGA to HDMI connector and wait a while to connect using viewer but at the end of it all, you should be successful.</p><p>Stay tuned for part II.</p><p><strong>Important links for further reading</strong></p><p><a href="https://www.realvnc.com/en/connect/docs/unix-start-stop.html#unix-start-stop">https://www.realvnc.com/en/connect/docs/unix-start-stop.html#unix-start-stop</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d246efe377d6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-research-lab-kampala/getting-an-object-detector-tracker-up-and-running-on-raspberry-pi-3-part-i-d246efe377d6">Getting an object detector/tracker up and running on Raspberry Pi 3 (Part I)</a> was originally published in <a href="https://medium.com/ai-research-lab-kampala">Makerere AI Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I (manually) converted a Computer Vision application from Python to Java]]></title>
            <link>https://medium.com/ai-research-lab-kampala/how-i-manually-converted-a-computer-vision-application-from-python-to-java-2cd4264b061b?source=rss----ca834513b05---4</link>
            <guid isPermaLink="false">https://medium.com/p/2cd4264b061b</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[computer-vision-app]]></category>
            <category><![CDATA[java-to-python]]></category>
            <dc:creator><![CDATA[Flavia Ninsiima]]></dc:creator>
            <pubDate>Wed, 29 Aug 2018 14:59:48 GMT</pubDate>
            <atom:updated>2018-08-29T14:59:47.976Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/560/1*7qatCLgen2mVKCXJjU3kkA.jpeg" /></figure><p>Now you might wonder from the title why someone would want to delve into such a tedious task! It was simply because at one point in time, we wanted to develop an android application for our computer vision application. The application was written in python and used the libraries O<a href="https://opencv.org/">penC</a>V, <a href="http://www.numpy.org/">numpy</a> and S<a href="https://www.scipy.org/">cipy</a>. The main goal of this application was to perform segmentation tasks to extract an object from an image. In this blog post, I will describe the steps I took and challenges faced while converting the computer vision application code base from Python to Java, for those who might find themselves faced with a similar ordeal. Before we go further into the porting of python to java, I provide a brief disclaimer!</p><p><strong>Disclaimer:</strong> I am neither a maestro at this nor do I claim to be one! I am simply a research student that had a computer vision application written in python and wanted to translate it to java for the development of an android app.</p><p>Now let’s examine the case at hand. We want to read an image, do some image processing operations with it, and hopefully return some meaningful result! <a href="https://opencv.org/">OpenCV</a> is a computer vision library and has bindings for Python, Java and C++. <a href="http://www.numpy.org/">Numpy</a> and <a href="https://www.scipy.org/">Scipy</a> are libraries for scientific computing used to do tasks for example performing numerical operations on images as array objects.</p><p><strong>Note:</strong> Inherently, an image read with OpenCV is a<a href="https://docs.opencv.org/2.4/doc/tutorials/core/mat_the_basic_image_container/mat_the_basic_image_container.html"> Mat object</a> however, most of us choose to access pixels in the image using numpy array operations.</p><p>Let’s begin by getting the obvious out of the way. Translating python statements to java. What are the challenges?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/560/1*wRAgeuQNRBl4sTMkMNw7yQ.jpeg" /></figure><ol><li>Always declare the variables. For most of us who are used to a ‘loosely typed’ language like python, we have to remember to first declare an object before we can initialize it. More so, we have to know the type of object. Sometimes, one line of code in python translates to 3 lines of code in java.</li></ol><pre>#python<br>a = 3</pre><pre>//java<br>int a = 3;</pre><p>2. Always end the line of code with a semi-colon. Sounds easy enough, right? <strong>Except it always easy to forget to add it!</strong></p><p>3. Watch out for the braces that encapsulate all the functions. Sounds easy too, right? Maybe so. But it’s not as if the compiler will point to the actual line where your missing curly brace should be.</p><p>4. Some of the Object dtypes naturally inherent in python might not have <strong>direct</strong> translations in java. This would probably necessitate you to search online for a suitable representation.</p><p>5. Functions in java must have an appropriate return type and this can be <em>void</em> or any other java type. However, what happens when you have to return multiple objects, all with different dtypes. While there are number of suggestions online for this, I settled for returning an array of type <em>Object</em>.</p><pre>#python<br>def do_something():<br>    return x, y, z</pre><pre>//java<br>static Object[] do_something(){<br>    Object[] obj1 = Object[]{x,y,z};<br>    return obj1;<br>}</pre><p>6. etc</p><p>Most likely if you ever get stuck and can’t debug the code, forums like S<a href="http://www.stackoverflow.com/">tackOverflow</a> are of great help!</p><p>Let’s move on with transporting our opencv operations from python to java. Luckily this was somewhat easy as opencv has implementations for both languages. But what are the challenges.</p><ol><li>Installing OpenCV for Java. The <a href="https://opencv-java-tutorials.readthedocs.io/en/latest/01-installing-opencv-for-java.html#introduction-to-opencv-for-java">documentation</a> clearly explains this. But what if you want to setup development environments such as IntelliJ, Ecllipse, NetBeans for using <em>OpenCV Java. </em>A simple search online should lead you to tutorials on how to do this. In my case I had NetBeans, and I used <a href="https://www.codeproject.com/Tips/717283/How-to-Use-OpenCV-with-Java-under-NetBeans-IDE">this</a> tutorial to add opencv to the IDE.</li><li>Importing opencv functionalities. Whereas in python it’s sufficient to type import cv it’s not the case in the java implementation. OpenCV Java subdivides the functionalities into different <a href="https://docs.opencv.org/java/3.1.0/">packages</a> and you must know which function resides under which package. This can usually be deduced from the OpenCV documentation.</li></ol><pre>import org.opencv.core.Mat;<br>import org.opencv.imgcodecs.Imgcodecs.CV_LOAD_IMAGE_COLOR;<br>import org.opencv.imgcodecs.Imgcodecs.imread;<br>import org.opencv.imgcodecs.Imgcodecs.imwrite;<br>import org.opencv.imgproc.Imgproc;</pre><p>3. Return values of functions in JavaCV are explicitly declared before use and are assigned “<em>inplace”</em>.</p><pre>&#39;&#39;&#39;converting and image to grayscale&#39;&#39;&#39;<br>#python<br>imgray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)</pre><pre>//java<br>Mat imggray = new Mat();<br>Imgproc.cvtColor(img, imggray, Imgproc.COLOR_BGR2GRAY);</pre><p>Let’s examine where we might have cases of numpy and scipy being used, and challenges arising.</p><ol><li>Basic numpy operations such as subtracting. Keep in mind that in javacv we load the image as a matrix not an array.</li></ol><pre>&#39;&#39;&#39; Calculate a mask for a region &#39;&#39;&#39;</pre><pre>#python<br>    imb1 = np.ones_like(imc)<br>    imbr = imb1 - imc<br>    imbf = np.logical_and(imb, imbr)</pre><pre>//java<br>    Mat imb1 = Mat.ones(imc.rows(), imc.cols(), CvType.CV_8UC1);<br>    Core.subtract(imb1, imc, imbr);        <br>    Core.bitwise_and(imb, imbr, imbf);</pre><p>2. Comparing images</p><pre>#python<br>Mat imgt1 = blurredImg &lt; 150;<br>Mat imgt2 =  blurredImg &gt; 25;<br>mask = np.bitwise_and(imgt1, imgt2);</pre><pre>//java<br>Mat imgt1 = new Mat(); Mat imgt2 = new Mat(); Mat mask = new Mat();<br>Core.compare(blurredImg, new Scalar(150), imgt1, Core.CMP_LT);<br>Core.compare(blurredImg, new Scalar(50), imgt2, Core.CMP_GT); <br>Core.bitwise_and(imgt1, imgt2, mask);</pre><p>3. Slicing an image</p><pre>&#39;&#39;&#39; Slice using bounding rect &#39;&#39;&#39;<br>#python<br>    mask = fullmask[rowstart:rowend,colstart:colend]</pre><pre>//java<br>    Mat mask = fullmask.rowRange(r1,r2).colRange(c1,c2);</pre><p>This is only a snapshot of what transpired. At one point in time I had to figure out how to load a python pickled object into java. I hope you enjoyed this post.</p><h4><strong>Important links for further reading:</strong></h4><p><a href="https://medium.com/onfido-tech/live-computer-vision-with-opencv-on-mobiles-f4e5ab15ad48">https://medium.com/onfido-tech/live-computer-vision-with-opencv-on-mobiles-f4e5ab15ad48</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2cd4264b061b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-research-lab-kampala/how-i-manually-converted-a-computer-vision-application-from-python-to-java-2cd4264b061b">How I (manually) converted a Computer Vision application from Python to Java</a> was originally published in <a href="https://medium.com/ai-research-lab-kampala">Makerere AI Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Re-training an object detection model on a custom dataset with Imageai, Clouderizer and Google…]]></title>
            <link>https://medium.com/ai-research-lab-kampala/re-training-an-object-detection-model-on-a-custom-dataset-with-imageai-clouderizer-and-google-c920c984f656?source=rss----ca834513b05---4</link>
            <guid isPermaLink="false">https://medium.com/p/c920c984f656</guid>
            <category><![CDATA[google-colab]]></category>
            <category><![CDATA[clouderizer]]></category>
            <category><![CDATA[imageai]]></category>
            <dc:creator><![CDATA[Flavia Ninsiima]]></dc:creator>
            <pubDate>Wed, 15 Aug 2018 08:31:30 GMT</pubDate>
            <atom:updated>2018-08-15T08:31:29.912Z</atom:updated>
            <content:encoded><![CDATA[<h4>Re-training an object detection model on a custom dataset with Imageai, Clouderizer and Google Colab</h4><p>Recently at a Data Science Africa conference(<a href="http://www.datascienceafrica.org/">dsa2018</a>), I was introduced to the amazing and wonderful Google Colaboratory — a service that runs on google drive and allows you to run your notebooks with free access to a Gpu.</p><p>Now I thought to myself, Great! Finally I can run all those pesky models that need intense computing power. Well… it turns out that getting stuff up and running on Colab for the first time wasn’t fun and exciting. What exactly wasn’t the fun part, figuring out how to reference my training data in the Colab notebook.</p><p>You see, recently I had come across a great python package called <a href="https://github.com/OlafenwaMoses/ImageAI">ImageAI</a>, so cool that it promises you can train a custom prediction model in 5 lines of code. Yeah, wow!!! How awesome is that! However, the re-training is based off the current models such as ResNet, InceptionV3. And due to the complexity of these models, it wasn’t such a great an experience re-training using a CPU.</p><p>So here I was, trying to figure out how to load my data with the Colab notebook. Now look here, yes there are ways to do this, definitely, but not so fun &amp; exciting. The point being, I simply didn’t anticipate the hassle of figuring out how to upload my training data into Google Drive and then reference it in my Colab notebook.</p><p>So anyhow fast forward, I decided to use whichever method I landed on first i.e. Drive FUSE wrapper. So there I was re-using a code referenced in so many tutorials, when lo and behold! an error pops up and I couldn’t for the life of me figure out what the heck the big deal was (since after all I had copied and pasted a code that had been re-used in all the searched for FUSE tutorials, ha!)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u_xFFJX14UmEe0_XSrnnkg.png" /></figure><p>So I gave up (this experience with Colab happened in the space of one afternoon), and I called a day a day! THE END.</p><p>And then there popped into my mail a Kaggle email titled (Access Free GPUs for Deep Learning) from a Mr. Jamie@kaggle. So hooray! I’m pumped, i’m ready to hit this thing, rock’n’roll. I headover to Kaggle, select Kernels, new Kernel type ‘Notebook’</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*igZkZAXlkExVfAwPk3dGqA.png" /></figure><p>I then try to figure out how to upload my data of course, but wait…. ah! You see the ImageAI module is a custom package… and while Kaggle allows you to install custom modules, this is “probably” if they appear in <a href="https://pypi.org/">PyPi</a>. Well… turns out this wasn’t the case for ImageAI, and I din’t want to have to first send a request to Kaggle to get it included by default (because this all happened in the space of one morning), and I wasn’t so patient.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7iVKNZKfGOljl-ytb4sLMQ.png" /><figcaption>Tap into the power of GPUs with Kaggle Kernels! Simply click the new “Enable GPU” checkbox on the Settings tab of your script or notebook</figcaption></figure><p>So there I am, when i see my partner-in-crime at the lab — Benji, training a dope, very dope R-CNN with the services of google cloud platform (because he is just so lucky that someone just happened to have some credits he could use for free). And there I was asking how I could tap into some of these credits, when lo and behold! he introduces me to Clouderizer. And guess what, I could tap into the GPU services that Colab was offering. Yippee! So he gives me a great, brief tutorial on how to get started, AND I START!</p><h4>Setting up Clouderizer and Google Colab</h4><p>Honestly, while it would be great to outline the following steps with pictures, the <a href="https://medium.com/@prakash_31206/clouderizer-integrates-with-google-drive-more-space-for-our-models-and-datasets-a842f23bca0a">Clouderizer blog</a> does it best!</p><p>In short we <strong>Create an account with Clouderizer</strong><em> and </em><strong>Link it with google-drive</strong>. At the end of this all we should have a <em>clouderizer</em> folder in our drive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*pQ8TM13s1joj5x89Oh9vmg.png" /></figure><p>Next <strong>Create a new project</strong> using <em>Local Machine</em> (if you wish to run it in your desktop) however there are options for <em>Amazon AWS</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CsfMJDaPJds2xyi3_dioEg.png" /></figure><p>Enable <strong>Remote Access</strong> with <em>terminal </em>and <em>jupyter notebook</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QU1tEHGqMAwTNXIJZh_FFg.png" /></figure><p>At the WORKSPACE stage, clouderizer will create default sub-directories <strong>code</strong> / <strong>data</strong> / <strong>out</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WLv-H5QqS4wL1hC6piulig.png" /></figure><p>This translates into the following in your google drive</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jH7TVGkVAxUIFYVcSGbf1A.png" /></figure><p>At the FINALIZE stage, you will be given two links: linux &amp; os x</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kWIET8qCDztEOJUwmJX2FQ.png" /></figure><h4>Getting the project up and running on drive</h4><ol><li>In google drive, start up a new colab notebook and enable GPU. See <a href="https://medium.com/deep-learning-turkey/google-colab-free-gpu-tutorial-e113627b9f5d">this</a> great tutorial on how to do so.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*OMJ4wZ8GYpeQYonAip61ng.png" /><figcaption>Borrowed from <a href="https://towardsdatascience.com/fast-ai-lesson-1-on-google-colab-free-gpu-d2af89f53604">https://towardsdatascience.com/fast-ai-lesson-1-on-google-colab-free-gpu-d2af89f53604</a></figcaption></figure><p>2. Copy and paste the linux link into the notebook and run as a command (run until you see a revolving stop sign)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nAdzzuaWvPSBeq-kJIvnOw.png" /></figure><p>3. On the clouderizer end, refresh the projects page to see if the project is running. Fire up your jupyter notebook and you will see a panel that enables you to explore your directories in drive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6oFA6FljZnp_a73VBhD9EA.png" /><figcaption>The jupyter and terminal buttons should become active after a while</figcaption></figure><p>4. Select Jupyter! Run your deep learning notebooks. VOILA! You can now <a href="https://github.com/OlafenwaMoses/ImageAI/blob/master/imageai/Prediction/CUSTOMTRAINING.md">retrain a network with ImageAI in 5 lines of code.</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EtMxOfdWCUo-M-M1XP3ikA.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c920c984f656" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-research-lab-kampala/re-training-an-object-detection-model-on-a-custom-dataset-with-imageai-clouderizer-and-google-c920c984f656">Re-training an object detection model on a custom dataset with Imageai, Clouderizer and Google…</a> was originally published in <a href="https://medium.com/ai-research-lab-kampala">Makerere AI Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AlexNet: A Brief Review]]></title>
            <link>https://medium.com/ai-research-lab-kampala/alexnet-a-brief-review-14979ce7cc84?source=rss----ca834513b05---4</link>
            <guid isPermaLink="false">https://medium.com/p/14979ce7cc84</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[convolutional-network]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Benjamin Akera]]></dc:creator>
            <pubDate>Fri, 10 Aug 2018 12:02:34 GMT</pubDate>
            <atom:updated>2020-05-22T14:47:10.812Z</atom:updated>
            <content:encoded><![CDATA[<p>AlexNet is a Convolutional Neural Network that rose to prominence when it won the Imagenet Large Scale Visual Recognition Challenge <a href="http://www.image-net.org/challenges/LSVRC/">(LSVRC)</a>, which is an annual challenge that evaluates algorithms for object detection and image classification at large scale (think of it as the World cup for image classification algorithms).</p><p>The ILSVRC evaluates the success of image classification solutions by using two important metrics, the top-5 and the top- 1 error. When given a set of N images, often called test images and mapped to a target class for each metric. The Top-1 Error is the percentage of the time the classifier did not give the correct class the highest score while the top-5 error is the percentage of the time that the classifier did not include the correct class among its top 5 guesses.</p><p>AlexNet received a top-5 error around 16% which was an extremely good result back in 2012. To put in context, the next best result trailed far behind (26.2%). When the dust settled deep learning became cool again and in the next few years, multiple teams would build CNN architectures that would beat human-level accuracy. The architecture used in the <a href="https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks">2012 paper</a> is popularly called AlexNet after the first author Alex Krizhevsky.</p><p>In this blog post, we will have a look at the details of the Alexnet architecture and try to re-implement it in Keras. Let&#39;s dive in!</p><h4><strong>Data</strong></h4><p>The model was trained on imageNet data which contains about 1.2 million images of 1000 categories.</p><h4>Image pre-processing</h4><p>Since ImageNet images are variable resolution, and the model presented in this paper requires fixed-size images, they scaled every image to 256x256 pixels. The scaling like so:</p><ol><li>Scale a possibly rectangular image so that the shorter side is 256 pixels.</li><li>Take the middle 256x256 patch as the input image.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/650/0*3yP4J15GU0S6gMG5" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2O0qEKwCZNfBPskR-RZuKQ.png" /><figcaption>preprocessing</figcaption></figure><h4>Data Augmentation</h4><p>This architecture used two forms of data augmentation.</p><ol><li>Translation and reflection. This consists of generating image translations and horizontal reflections. This scheme helped increase the size of the data by a scale of 2048 without which the network would have suffered from substantial overfitting.</li><li>The other way they augmented the dataset involved perturbing the R,G,B values of each input image by a scaled version of the principal components (in RGB space) across the whole training set. This scheme helped reduce the top-1 error rate by over 1%.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wYG9AO1Kh30tMj6F93sP-A.png" /></figure><h4>Architecture</h4><p>AlexNet is made up of eight trainable layers, five convolution layers and three fully connected layers. All of the trainable layers are followed by a ReLu activation function except for the last fully connected layer where a softmax function is used. The architecture also consists of non-trainable layers: Three pooling layers, two normalization layers and one dropout layer (used to reduce overfitting).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*arJJYgK-_7VcuKKX1TCKMA.png" /><figcaption>AlexNet architecture</figcaption></figure><h4>Choice of Non-Linearity.</h4><p>The authors chose to use the Rectified Linear Unit (ReLU) function. They saw that deep convolutional neural nets with ReLUs trained several times faster than their equivalents with <strong>tanh </strong>units. When pitted against the tanh activation with no other changes, they were able to train their model to 25% error rate on the training set 6X faster with the ReLU activation.</p><h4>Training</h4><p><a href="https://en.wikipedia.org/wiki/Stochastic_gradient_descent">Stochastic Gradient Descent</a> with learning rate 0.01, momentum 0.9 and weight decay 0.0005 is used. The training is done on two GPUs (GTX 580) for parallelism, and the setup is quite interesting. The GPUs used each have 3GB memory. The network is split into halves, as can be seen in the model description figure, across the two GPUs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/631/1*cZvLDHIMLdWCrilgDKPQQw.png" /><figcaption>Training details</figcaption></figure><h4>Implementation</h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/03f93af6d097ce5b7c41f0c244aa2acc/href">https://medium.com/media/03f93af6d097ce5b7c41f0c244aa2acc/href</a></iframe><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/44275d7f94e7d7519bd9f45141abea7d/href">https://medium.com/media/44275d7f94e7d7519bd9f45141abea7d/href</a></iframe><h4>Conclusion.</h4><p>This work was the first of its kind to have trained deep convolutional networks on GPUs to achieve impressive results on the ImageNet dataset for image classification. I hope you got to learn something.</p><p>Later.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=14979ce7cc84" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-research-lab-kampala/alexnet-a-brief-review-14979ce7cc84">AlexNet: A Brief Review</a> was originally published in <a href="https://medium.com/ai-research-lab-kampala">Makerere AI Lab</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>