<?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[Stories by Latent Cybernetics and Praxis on Medium]]></title>
        <description><![CDATA[Stories by Latent Cybernetics and Praxis on Medium]]></description>
        <link>https://medium.com/@philippp?source=rss-39fe797df35------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*VIxIJRSBKLw2lpb4.</url>
            <title>Stories by Latent Cybernetics and Praxis on Medium</title>
            <link>https://medium.com/@philippp?source=rss-39fe797df35------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 16:20:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@philippp/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[DNS for your local network]]></title>
            <link>https://medium.com/@philippp/dns-for-your-local-network-8dec153c9558?source=rss-39fe797df35------2</link>
            <guid isPermaLink="false">https://medium.com/p/8dec153c9558</guid>
            <category><![CDATA[networking-tips]]></category>
            <category><![CDATA[dns-servers]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[dns]]></category>
            <category><![CDATA[home-improvement]]></category>
            <dc:creator><![CDATA[Latent Cybernetics and Praxis]]></dc:creator>
            <pubDate>Wed, 26 Apr 2023 19:16:33 GMT</pubDate>
            <atom:updated>2023-04-26T19:16:33.766Z</atom:updated>
            <content:encoded><![CDATA[<p>Does your local network have a number of servers with static local IPs? <br>Have you considered setting up DNS, but couldn’t get it to work reliably?</p><p>Me too, and that’s why I wrote this guide!</p><ol><li>Setting up your DNS server</li><li>Configuring your DHCP server to forward the DNS server’s address</li><li>Troubleshooting resolution issues</li></ol><h3>Setting up your DNS server</h3><p>Many options and opinions; for my use case of “a small number of local vanity domains” <a href="https://thekelleys.org.uk/dnsmasq/doc.html">dnsmasq</a> seemed like a good fit. There are many good setup guides; in the end you will specify your vanity domains in the /etc/hosts file of the machine running dnsmasq.</p><h3>Configuring your DHCP server</h3><p>It’s likely that your router is also acting as a DHCP server. Look around the settings for additional DNS servers and specify the IP of the machine running dnsmasq (your local nameserver), or add this IP to the list of nameservers to forward to DHCP clients.</p><h3>Troubleshooting Debian and Ubuntu</h3><p>Modern Debian and Ubuntu uses <em>resolved, </em>which is a local DNS server running on 127.0.0.53. The <em>resolvectl</em> command reports the DNS servers that requests to 127.0.0.53 are forwarded to, which SHOULD include your local nameserver. If you do not see your local DNS server in the list, check our DHCP server (likely on your router) and make sure that the nameserver is specified and changes have been applied.</p><p>If your nameserver is properly specified but domains don’t resolve on a network peer, try running <em>dig @local.nameserver.ip.address custom.domain </em>from the network peer to ensure the nameserver is reachable.</p><p>If the nameservice is reachable and returns an A (or other relevant) record, make sure your custom domain includes a (fake) suffix, such as .home.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8dec153c9558" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Converting GeoJSON to STL for 3D printing]]></title>
            <link>https://medium.com/@philippp/converting-geojson-to-stl-for-3d-printing-1ad56e4d3058?source=rss-39fe797df35------2</link>
            <guid isPermaLink="false">https://medium.com/p/1ad56e4d3058</guid>
            <category><![CDATA[3d-modeling]]></category>
            <category><![CDATA[gis]]></category>
            <category><![CDATA[makers]]></category>
            <category><![CDATA[3d-printing]]></category>
            <dc:creator><![CDATA[Latent Cybernetics and Praxis]]></dc:creator>
            <pubDate>Fri, 06 Jan 2023 19:17:57 GMT</pubDate>
            <atom:updated>2023-01-09T05:21:15.986Z</atom:updated>
            <content:encoded><![CDATA[<h3>My 7 Hills: GeoJSON to STL</h3><p>tl;dr: We reconstruct an STL mesh from GeoJSON data by extracting the GeoJSON shape coordinates into a point cloud and applying Poisson reconstruction using Open3d.</p><p>Deeply unsatisfied with my first attempt at printing GeoJSON hills, I came up with a three-step process that produces high-fidelity reproductions using only open source software. This process has three steps:</p><ol><li>Slice the large geojson map using <a href="https://geopandas.org/">Geopandas</a>.</li><li>Create an XYZ point cloud of the geojson geometry (leveraging <a href="https://shapely.readthedocs.io/">Shapely</a> and Geopandas).</li><li>Build a TriangleMesh for the XYZ point cloud using <a href="http://www.open3d.org/docs/release/">Open3D</a>.</li><li>Extrude a solid from this mesh using <a href="https://www.blender.org/">Blender</a>.</li></ol><p>For this iteration, we’ll try to model the cliffside including Telegraph hill. If you want to try this yourself, here’s <a href="https://github.com/philippp/geojson2stl">my script</a> and the <a href="https://data.sfgov.org/Energy-and-Environment/Elevation-Contours/rnbg-2qxw">SF GeoJSON</a> data.</p><figure><img alt="Google Maps screenshot of Telegraph Hill" src="https://cdn-images-1.medium.com/max/1024/1*HzolkW2uOKoKaAVJBQIG2A.png" /><figcaption>Telegraph Hill is an iconic cliff-sided elevation topped by Coit Tower.</figcaption></figure><h3>Step 1: Slicing the earth</h3><p><a href="https://geopandas.org/en/stable/">GeoPandas</a> extends Pandas dataframes with geodata and spacial primitives. It allows us to clip shapes to our <em>Longitude, Latitude</em> designated window of interest, convert coordinates to <a href="https://en.wikipedia.org/wiki/Universal_Transverse_Mercator_coordinate_system"><em>UTM X and Y meters</em></a>, and more.</p><p>First, let’s slice the SF geojson map to only include features (and aspects of features) within the coordinates of Telegraph Hill:</p><pre>#!/usr/bin/env python3<br>import geopandas<br>from shapely.geometry import Polygon<br><br>outname = &quot;telegraph_hill&quot;<br>y1, x1 = (37.8053307084282, -122.40853235179131)<br>y2, x2 = (37.79953896610929, -122.40128101638189)<br><br>gdf = geopandas.read_file(&#39;geojson/sf.geojson&#39;)<br>polygon = Polygon([(x1, y1), (x1, y2), (x2, y2), (x2, y1), (x1, y1)])<br>gdf_clipped = gdf.clip(polygon)<br>gdf_clipped.to_file(f&quot;geojson/{outname}.geojson&quot;, driver=&#39;GeoJSON&#39;)<br>gdf_clipped = gdf_clipped.explode(index_parts=True)  #Line Segments only!</pre><h3>Step 2: Generating a Point Cloud from GeoJSON</h3><p>Step 1 gave us a “GeoDataFrame” of shapes within our desired coordinates. We now need to convert these relevant shapes to a point cloud, and then into a triangle mesh for 3d printing.</p><p>We will continue to use geopandas to translate the coordinates from <em>Long, Lat</em> to <em>X, Y meters from origin</em>, and join these with the Z data from our source data’s ‘elevation’ property. These datapoints then initialize an Open3d point cloud.</p><pre>import open3d as open3d<br><br>CONV_FT_TO_M = 0.3048  # SFData provides elevation in feet :(<br><br>def compute_gdf_pointcloud(gdf: geopandas.GeoDataFrame, z_scale: float) -&gt; o3d.geometry.PointCloud:<br>    &quot;&quot;&quot;Compute the PointCloud of the GeoDataFrame.&quot;&quot;&quot;<br>    min_x, min_y = (99999999,99999999)<br>    min_elevation = min([e for e in gdf[&#39;elevation&#39;]])<br>    <br>    for f in gdf[&#39;geometry&#39;]:<br>        for c in f.coords:<br>            if c[0] &lt; min_x:<br>                min_x = c[0]<br>            if c[1] &lt; min_y:<br>                min_y = c[1]<br><br>    logging.debug(f&quot;min_x={min_x} min_y={min_y} min_elevation={min_elevation}&quot;)<br>    gdf[&#39;flat_coordinates&#39;] = gdf[&#39;geometry&#39;].combine(<br>        gdf[&#39;elevation&#39;],<br>        (lambda g, e: [(float(c[0] - min_x),<br>                        float(c[1] - min_y),<br>                        float(e) * CONV_FT_TO_M * z_scale) for c in g.coords]))<br>    # Add these coordinates to an Open3d point cloud.<br>    pcd = o3d.geometry.PointCloud()<br>    pcd.points = o3d.utility.Vector3dVector()<br>    for coordinate_list in gdf[&#39;flat_coordinates&#39;]:<br>        pcd.points.extend(o3d.utility.Vector3dVector(coordinate_list))<br>    # Estimate normals.<br>    pcd.estimate_normals(search_param=o3d.geometry.KDTreeSearchParamHybrid(radius=0.01, max_nn=30))<br>    pcd.orient_normals_to_align_with_direction()<br>    return pcd</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1NCnAniazgDfGx_bRN8g9A.png" /></figure><h3>Step 3: Build a TriangleMesh for the XYZ point cloud</h3><p>We now attempt Poisson mesh reconstruction from our point cloud with normals. The Poisson mesh reconstruction also returns density data for each inferred triangle, which allows us to trim extraneously inferred areas if necessary. (It also the only interesting coloration we have)</p><pre>import open3d as o3d<br>import numpy as np<br>import matplotlib.pyplot as plt<br><br>def compute_poisson_mesh(pcd: o3d.geometry.PointCloud, depth: int) <br>   -&gt; o3d.geometry.TriangleMesh:<br>    &quot;&quot;&quot;Compute the mesh of the point cloud.<br>    depth:    The depth of the octree used for the surface reconstruction determines resolution of the resulting triangle mesh.<br>    &quot;&quot;&quot;<br>    poisson_mesh, densities = o3d.geometry.TriangleMesh.create_from_point_cloud_poisson(<br>        pcd, depth=depth, width=0, scale=1.1, linear_fit=True)<br>    # Color the supporting point density.<br>    densities = np.asarray(densities)<br>    graph_rate = (densities - densities.min()) / (densities.max() - densities.min())<br>    density_colors = plt.get_cmap(&#39;plasma&#39;)(graph_rate)<br>    density_colors = density_colors[:, :3]<br>    poisson_mesh.vertex_colors = o3d.utility.Vector3dVector(density_colors)<br>    # Trim any excess surface from the Poisson reconstruction.<br>    bbox = pcd.get_axis_aligned_bounding_box()<br>    p_mesh_crop = poisson_mesh.crop(bbox)<br>    p_mesh_crop.compute_triangle_normals()<br>    return p_mesh_crop<br></pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4Rh8yh8-OUzK5DkL28kq8w.png" /></figure><p>Altogether not bad! Inferred from the point density, the color of the mesh indicates how many points (from our pointcloud) indicate confidence in the given polygon. Since our process benefits from a rectangular mesh (see the next step), we don’t trim any excess area.</p><h3>Step 4: Fixing and printing</h3><p>We need to fix a few holes in our geometry and optionally create an efficient solid (mountains should be hollow). To fix holes, import your STL into Blender and select the vertices surrounding the hole in edit mode and hit “alt-F” to fill the hole with triangles.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/564/1*om86SjS3_JFu1GGZgT8QIg.png" /></figure><p><em>A generalized solution for STL hole patching is left as an exercise for the reader.</em></p><p>We should now be able to extrude our mesh and use the Bisect tool to preserve the upper part. Check for non-manifold faces along the way, and recompute normals before exporting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KBtyTxd4Af8-jaw44u9zKQ.jpeg" /></figure><p>This is much nicer than what I was initially expecting! I hope this helps you in your journey, and happy hacking.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1ad56e4d3058" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My 7 hills: GeoJSON and 3D printing]]></title>
            <link>https://medium.com/@philippp/my-7-hills-geojson-and-3d-printing-5745de0aad1a?source=rss-39fe797df35------2</link>
            <guid isPermaLink="false">https://medium.com/p/5745de0aad1a</guid>
            <category><![CDATA[san-francisco]]></category>
            <category><![CDATA[geojson]]></category>
            <category><![CDATA[geospatial]]></category>
            <category><![CDATA[3d-printing]]></category>
            <category><![CDATA[3d-modeling]]></category>
            <dc:creator><![CDATA[Latent Cybernetics and Praxis]]></dc:creator>
            <pubDate>Tue, 03 Jan 2023 08:22:43 GMT</pubDate>
            <atom:updated>2023-01-03T21:16:53.183Z</atom:updated>
            <content:encoded><![CDATA[<p><em>tl;dr: I converted Lon, Lat polygon data into a 3d print, and you can, too!</em></p><p>I adore San Francisco’s hills and would love to better incorporate their geometry into ongoing art projects. While I was unable to find high-resolution STL files of either San Francisco’s topography or individual hills, the SFData site has an <a href="https://data.sfgov.org/Energy-and-Environment/Elevation-Contours/rnbg-2qxw">elevation map</a> with five feet of resolution.</p><figure><img alt="Screenshot of the SFData site showing the contour lines of the GEOJson elevation map." src="https://cdn-images-1.medium.com/max/1024/1*c5fmlEeOD35jk_Bl3gV10w.png" /><figcaption>The contour lines surrounding Telegraph Hill, as seen on data.sfgov.org</figcaption></figure><p>This elevation map is a geojson file of Polygons and lines, with each point indicated as (Longitude, Latitude) coordinates. Regrettably, the conversion of GeoJSON to STLs is challenging to generalize: First, GeoJSON does not specify the “normals” (i.e. which side is “inside” and which side is “outside”), also lines do not necessarily form closed surfaces.</p><p>Not to panic — perhaps we can start with the simple geometry of hills formed by closed polygons and later figure out how to make use of line data.</p><h3>Step 1: Programmatic Hill Identification &amp; SVG export</h3><p>Let’s find hills that rise on all sides (vs. flatten into a higher plateau):</p><ol><li>We select all of the polygons within a (Lon, Lat) window</li><li>Sort polygons by z-index / vertical layer, and iterate from the highest to lowest z-index / layer:</li><li>If a point on polygon A at a given z-index Z is on polygon B at z-index Z-1, the polygon B considers polygon A a “leaf.” (I.e. Leaf polygons sit on top of other polygons)</li><li>Assuming no overhangs, if any point on polygon A at z-index Z is not on any polygon at z-index Z-1, the current polygon is a “root”</li></ol><p>We assume no overhangs, and that the hill rises slowly enough such that every z-layer has new polygon coordinates. This process gives us a number of root nodes with leaves (which in turn may have other leaves), which constitute our hills.</p><p>Let’s render these and sanity check our results by exporting to SVG. SVG is written as plain text XML, and a polygon is simply a list of X,Y points in which the first point is equal to the last point.</p><figure><img alt="Screenshot showing excerpt from an SVG file." src="https://cdn-images-1.medium.com/max/728/1*NY2bOKh67VERgsj-uwXRwA.png" /><figcaption>Excerpt from SVG file describing Telegraph Hill’s topography.</figcaption></figure><figure><img alt="Rendered topographic lines of Telegraph Hill." src="https://cdn-images-1.medium.com/max/1024/1*-rvHs1ZmkuCLm97vU60ulQ.png" /><figcaption>Topographic lines of Telegraph Hill</figcaption></figure><p>Perhaps the simplest way to print this in 3D is to turn each polygon into a “wall” of fixed height, and stack these walls vertically.</p><p>Fortunately, the SFData GeoJSON file specifies the layer height in feet, and the Haversine formula reveals the real-world length and width from the Longitude and Latitude of our Polygon data. Once these are converted into the same units (offset from a reference point in meters), we can scale every point of every line to X, Y, Z coordinates.</p><h3>Step 2: Export to STL and Print</h3><p>We now have concentric polygons of varying heights (i.e. Z-values). I would like to turn each of these into walls by drawing vertical planar shapes. We can semi-trivially do this by hand:</p><p><strong>STL Files </strong>are just lists of triangles called “facets,” each of which has three vertices with X,Y,Z coordinates.</p><pre>facet normal ni nj nk<br>    outer loop<br>        vertex v1x v1y v1z<br>        vertex v2x v2y v2z<br>        vertex v3x v3y v3z<br>    endloop<br>endfacet</pre><p>In order to create a vertical “wall” between two polygon points, we draw two complementary triangles — one pointing to the top, and one pointing to the bottom. The surface normal can be computed from the cross product of the triangle (right hand rule!)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sUsAP7PDhrZTdxmmecV8pg.png" /><figcaption>The top of Telegraph Hill</figcaption></figure><p>The astute reader may note that the polygon walls do not form a closed surface — meaning, there is no “top” to the mountain. Thankfully, my slicer is forgiving in this regard and Telegraph Hill printed reasonably well.</p><figure><img alt="Photograph of the 3D printed solid of Telegraph Hill" src="https://cdn-images-1.medium.com/max/1024/1*_tD-RuNvhcdlPbFjHJZbog.jpeg" /><figcaption>Telegraph Hill printed using Soy-based resin.</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5745de0aad1a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[3D print your own cookie cutter!]]></title>
            <link>https://medium.com/@philippp/make-your-own-cookie-cutter-b50424b66d65?source=rss-39fe797df35------2</link>
            <guid isPermaLink="false">https://medium.com/p/b50424b66d65</guid>
            <category><![CDATA[cookies]]></category>
            <category><![CDATA[3d-printing]]></category>
            <category><![CDATA[art]]></category>
            <category><![CDATA[makers]]></category>
            <category><![CDATA[baking]]></category>
            <dc:creator><![CDATA[Latent Cybernetics and Praxis]]></dc:creator>
            <pubDate>Fri, 30 Dec 2022 21:44:21 GMT</pubDate>
            <atom:updated>2022-12-30T21:48:46.779Z</atom:updated>
            <content:encoded><![CDATA[<p>What’s better during the Holiday season than fresh-baked cookies? Fresh backed cookies in shapes and designs of your own making, of course! Using only free and open-source software, you too can make your own cookie cutters — here’s how.</p><h3>Step 1: Choose a vector graphic to imprint</h3><p>First, we want to choose the shape we want to impress upon our cookie. I’m doing rectangular cookies imprinted with an outline of the SF bay area, but you could do any shape and print pattern. I found an annotated map of the bay area at <a href="https://www.researchgate.net/figure/Map-of-San-Francisco-Bay-California-indicating-key-locations-mentioned-in-the-text_fig3_282238231">https://www.researchgate.net/figure/Map-of-San-Francisco-Bay-California-indicating-key-locations-mentioned-in-the-text_fig3_282238231</a>, loaded it into Gimp and cleaned it up a bit: Removed the text labels, evened out the line thickness, and removed some details south of the San Mateo bridge. Inkscape allowed me to convert the bitmap it into a scalable vector graphic (SVG).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/738/1*W9-uORIZEMhzmlxyqmnTDg.png" /></figure><h3>Step 2: Model your cookie cutter</h3><p>I imported the vector graphic into Blender and scaled the path to 7cm x 9cm before converting it into a mesh in “object mode”. The “extrude” tool then extends the line art about 1cm along the z-axis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*st4UJGDDRZ4uYv0CZvNYJw.png" /></figure><p>Finally, place cuboids around the perimeter of the extruded outline. In the end, you will want ~5 mm of space between the edge of your rim and the edge of your line art. Depending on how well-connected your design is, you may want to add a few structural supports at the top.</p><p>Finally, you will want to use edit mode to create an edge by dragging the inner lip of your rim further away from the line art, and bringing the edges to a point at the corners.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/903/1*2lrASyXtI0R56tipHf_yTA.png" /></figure><p>Now join all of the elements into one object, export as a .STL file and into the slicer of your choice! I printed this once without supports and with a backing grid (top picture), but had a hard time removing it from my base plate. I opted to print it with supports (and without the backing grid) afterwards (see bottom picture).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cy8Lj2y_X_dpxYYszNa0OQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kxObpfj6ceNsDeCO3_fm3A.png" /></figure><h3>Step 3: Print, Cure, Test, Iterate!</h3><p>After a quick print + cure and pruning the support we do a first test print.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cXVRGDgIWfpopRYE" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/943/1*VxhNl2cyTUaRVnsk1h-X9g.jpeg" /></figure><p>The imprint is true-to-form, and even fine details are persisted. If I try this again, I may want to iterate on a few aspects:</p><ol><li>Since I printed the shape without an internal backing, we lost Angel Island. I really like Angel Island, so I should add an internal support for it.</li><li>The line thickness feels ok for San Francisco, but too thick for Alameda.</li><li>I’d like to try rendering the bridges as two parallel lines to set them apart visually.</li></ol><h3>Step 4: Cookies!</h3><p>I attempted some sugar cookies with the above bay area outline, as well as a San Francisco neighborhood map. Lessons learned:</p><p>If you press the decorative lines in too deep, the cookie dough will close them back up when expanding. Instead, try to keep the impression superficial, and it will persist through the rising process.</p><p>The dough loves getting stuck in the cutter, and getting the dough out will distort the shapes. It might be worth trying to oil the cookie cutter up before use in the future.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LZRZ5zYS_t0SK-nUWijXWQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b50424b66d65" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Distilling the fog of war]]></title>
            <link>https://medium.com/@philippp/distilling-the-fog-of-war-363f1289c189?source=rss-39fe797df35------2</link>
            <guid isPermaLink="false">https://medium.com/p/363f1289c189</guid>
            <dc:creator><![CDATA[Latent Cybernetics and Praxis]]></dc:creator>
            <pubDate>Sun, 27 Feb 2022 17:14:52 GMT</pubDate>
            <atom:updated>2022-02-27T17:19:49.532Z</atom:updated>
            <content:encoded><![CDATA[<p>As the world watches the suffering of the Ukrainian people and madness of Putin in horror, many of us are struggling to understand what is happening on the ground hour by hour, and day by day. Much information is posted in Telegram channels, which presents two challenges:</p><ol><li>Switching across a dozen channels, some with far more frequent updates than others, is inefficient and will miss information.</li><li>Those illiterate in Russian and Ukrainian are forced to use the web UI to get translations, but can only see some media in the native app.</li></ol><h3>Telethon and aggregation</h3><p>Telethon is a pure-python3 Telegram client API that allows you to both preview and join groups as you can on all other clients. You will need to register an app using your Telegram account and the <a href="https://my.telegram.org/apps">https://my.telegram.org/apps</a> service, noting the api_id and api_hash you will be using.</p><p>You can then create a client that simply “previews” the groups (using GetHistoryRequest), fetching all recent messages and media posted in the groups. Your client can then save the media (jpgs, mp4s) and render the messages and media embeds into an HTML file.</p><p>This way, you can pull up the HTML file with a local browser, which will load all embedded media and offer right-click translate abilities.</p><h3>So why don’t you build it?</h3><p>I do have this running locally, but it’s a LOT of media to serve to the broad public. If by any chance you are an investigative journalist or have other relevant needs though, hit me up and maybe I can help you out. ❤</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*-rrZYNSXRbfMQRQIKcCRnQ.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=363f1289c189" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Look ma, no Arduino! (μControl 101)]]></title>
            <link>https://medium.com/@philippp/look-ma-no-arduino-%CE%BCcontrol-101-37f9c5cfcd8f?source=rss-39fe797df35------2</link>
            <guid isPermaLink="false">https://medium.com/p/37f9c5cfcd8f</guid>
            <dc:creator><![CDATA[Latent Cybernetics and Praxis]]></dc:creator>
            <pubDate>Sat, 26 Feb 2022 20:23:47 GMT</pubDate>
            <atom:updated>2022-02-26T20:23:47.647Z</atom:updated>
            <content:encoded><![CDATA[<p>After playing with Arduino for a few projects and watching <a href="https://www.youtube.com/watch?v=ww_71dgkgW8&amp;t=522s">a motivational video</a> by way of a <a href="https://hak5.org/products/o-mg-demonseed-edu">demonseed edu kit</a>, I wanted to return to a beginner project (a remote controlled car) without depending on the affordances of an Arduino-like board.</p><p>I decided to use an ATTiny85 providing a 16.5MHz clockspeed, 5 i/o signal pins and allowing 8192 bytes of program instructions, 512 bytes of eeprom (static) memory and 2096 bytes of flash memory.</p><h3>Bootstrapping a microprocessor</h3><p>Getting instructions onto the microprocessor initially requires using a dongle (which uses another microprocessor!) to flash a bootloader onto the microprocessor. These are the instructions the processor runs when powered on. Following the instructions provided by <a href="https://twitter.com/_MG_">@MG</a>, I flashed a bootloader that allowed me to directly program the chip via USB. After flashing the bootloader and installing “Digispark” Arduino Board support, I’m able to upload code from the Arduino IDE using the following rigging:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cvwHMhXQc1CEQitAs3Qsuw.jpeg" /></figure><h3>Driving the motors</h3><p>My car has three wheels, two of which are powered and must be able to independently go backwards and forwards in order for the vehicle to turn. I decided to use an H-bridge in a DIP package, instead of going down the rabbit hole of building one from scratch. At first, the program would hang whenever the motors started, as they were drawing all available power from the battery. This was fixed by pairing each of the ICs with a sizeable capacitor.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Gj0I7k2pcSF37HUdRxDdXA.jpeg" /><figcaption>The top two ICs are L911DH (H-bridges) and the ATTiny85 is at the bottom.</figcaption></figure><h3>Remote controlling and Debugging</h3><p>I used a Roku remote and an IR sensor on the car to relay controls to the microprocessor. The IR sensor connects to pin 2 on the ATTiny, and will read high or low as the signal pulses. The program monitors for IR pulses, and attempts to match against known sequences stored in flash memory.</p><p><em>What are the sequences though?</em> It depends on the period between each sample, which is variable with what else is being done in our run loop. I ended up adding a UART driver to the program to report the observed values via serial. Since I only have 5 signal pins, I added a “debug” mode that uses the pins that usually inform the left motor correction for RX/TX. In non-debug mode, we do the same work (so that the signal matches), but set RX/TX to a non-existing pin.</p><h3>Demonstration</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FviGV4wyx5wg%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DviGV4wyx5wg&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FviGV4wyx5wg%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/ce48e40b5578b6d64af89ec7e96f6117/href">https://medium.com/media/ce48e40b5578b6d64af89ec7e96f6117/href</a></iframe><h3>Learnings</h3><p>I’m glad to have had a chance to play without the scaffolding of an Arduino-like platform. If I were to re-build this, I would add the USB reprogramming circuit onto the car itself to avoid having to re-seat the IC for every programming cycle. Also, I ended up frying a two ATTinys by plugging them in upside down — hardwiring the USB connections would prevent this.</p><p>The shortage of pins and inability to debug with full device functionality is also something I’d avoid — having an interactive debugger is clutch, especially as a beginner. The default failure state is “microcontroller is not doing anything,” whether that’s from using too much memory, a drop in input voltage, divide by zero or hardware failure. If you have been softened by rich stack traces and interactive debuggers, this will require some adjustment.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=37f9c5cfcd8f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Deprecated Genie]]></title>
            <link>https://medium.com/@philippp/the-deprecated-genie-e19cf4f0f210?source=rss-39fe797df35------2</link>
            <guid isPermaLink="false">https://medium.com/p/e19cf4f0f210</guid>
            <category><![CDATA[home-improvement]]></category>
            <category><![CDATA[wireless]]></category>
            <category><![CDATA[electronics]]></category>
            <dc:creator><![CDATA[Latent Cybernetics and Praxis]]></dc:creator>
            <pubDate>Sun, 19 Sep 2021 18:31:39 GMT</pubDate>
            <atom:updated>2022-12-19T16:20:29.535Z</atom:updated>
            <content:encoded><![CDATA[<p>Buying a house includes the acquisition of legacy systems. In our case, our garage is gated by an aging Genie garage door opener. There is a lively second market for the (deprecated) remotes, some of which sell for over $100 on Amazon — to compare, a new garage door opener including the motor <a href="https://www.amazon.com/Genie-Garage-Opener-Battery-Backup/dp/B079H23418/ref=asc_df_B079H23418/?tag=hyprod-20&amp;linkCode=df0&amp;hvadid=344004050138&amp;hvpos=&amp;hvnetw=g&amp;hvrand=4655495990412705420&amp;hvpone=&amp;hvptwo=&amp;hvqmt=&amp;hvdev=c&amp;hvdvcmdl=&amp;hvlocint=&amp;hvlocphy=9061254&amp;hvtargid=pla-780980562131&amp;psc=1&amp;tag=&amp;ref=&amp;adgrpid=69357499575&amp;hvpone=&amp;hvptwo=&amp;hvadid=344004050138&amp;hvpos=&amp;hvnetw=g&amp;hvrand=4655495990412705420&amp;hvqmt=&amp;hvdev=c&amp;hvdvcmdl=&amp;hvlocint=&amp;hvlocphy=9061254&amp;hvtargid=pla-780980562131">costs less than $200</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/561/1*oMG5Ooz8xHMSVzldcsYUnw.png" /><figcaption>So you want to get into your garage, eh?</figcaption></figure><p>Figuring myself smarter than the average bear, I opted for a $20 “compatible” device (the HEDDOLF G220). Following the instructions, I matched the DIP switches and… nothing. Not excited to shell out $100+ for a vintage remote, I decided to root out the issue. What’s being sent by the original remote, and how is the new remote different? And, if I’m going to have to guess at settings, how can I know whether I’m getting “warmer” or “colder?”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*cWgtlFU8iKSIBIcGyOM3Cw.gif" /><figcaption>Another bright mind excited about the world of radio telecommunications!</figcaption></figure><p>$25 will buy you a <a href="https://www.rtl-sdr.com/buy-rtl-sdr-dvb-t-dongles/">RTL-SDR dongle</a>, allowing you to “listen in” to frequencies between 500KHz and 1GHz. We’ll use a tool called <a href="https://github.com/jopohl/urh">Universal Radio Hacker</a> to both identify the frequencies of interest and analyze the waveforms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/943/1*eIxkbSNwK50UCSXhEh7DdQ.png" /><figcaption>The File menu links to both the Spectrum Analyzer and Record functionality.</figcaption></figure><p>We first identify the exact broadcast frequency by centering the spectrum analyzer on 390MHz, the frequency advertised on the back of the new remote.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lcYwueZEvasFM_bX6zcNVw.png" /><figcaption>Frequency Spectogram, showing a clear peak at 390.1MHz — close enough.</figcaption></figure><p>We then record a few seconds of this signal, which will allow us to identify what ones and zeroes are being sent over the air.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b9eRTXxkJJ8i_vQKVXJdMw.png" /></figure><p>We repeat the same for the original manufacturer’s remote, and save both captures in distinct files. We now switch over to the “Generator” tab in Universal Radio Hacker and enter the Modulation dialogue, to attempt to reconstruct the original signal.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vk3M3b-_tnTfC_g6Eo8qGQ.png" /><figcaption>Click on the “Generator” tab and then the “Edit…” button next to Modulation.</figcaption></figure><p>Here comes the fun part. We will attempt to generate a signal that is equivalent to the signal we recorded (for the original remote, and then the new remote). The “Modulation” interface does this by taking a <em>Carrier Frequency </em>and turning it on and off (i.e. sending 1s and 0s) in particular sequence. The fancy name for this is “Amplitude Shift Keying (ASK)” (remember this for future Radio enthusiast cocktail parties).</p><p>The Modulation interface allows you to preview the combination of a given carrier frequency and a sequence of 1s and 0s, and tweak these until they match the signal that you previously recorded. The process is simple, but the UI is not, so regard the following:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d6W5Bf7SViRz8RwpMWF2pQ.png" /></figure><ol><li>Drag and drop the recorded waveform into the bottom-right container. If you recorded more than a few seconds of signal, prepare to wait (or re-record). Zoom out until you see one period / full cycle of the signal that’s being sent.</li><li>The carrier frequency should be auto-detectable from the original signal. Try your luck!</li><li>Now the fun part. Enter 0s and 1s to define the data mask, and keep an eye on the “Modulation” preview. Once it looks close (in terms of on-off segments — the phase offset does not matter), you are good.</li><li>Write down the Carrier frequency and Data (raw bits), and exit.</li></ol><p>You can do this with both remotes to understand how they differ, and to try to get the new remote to match the old one.</p><p><strong>Bonus:</strong> If you have a radio that allows for transmission in addition to reception, you can click “Send Data” to broadcast the generated signal. If your garage door opens, you got it right. (Please abide by local laws and regulations regarding radio broadcasting.)</p><p>In the end, the new remote ended up working. For anyone who found this by Googling “Genie 912 remote” and “HEDDOLF G220” or similar, here’s what tripped me up: The 9/12 pin selector on the HEDDOLF remote is flipped — so if your 9/12 selector on the 912 is up, the HEDDOLF one must be down. The video and image below were most helpful:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FwtU8GL4oS1Y&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DwtU8GL4oS1Y&amp;image=http%3A%2F%2Fi.ytimg.com%2Fvi%2FwtU8GL4oS1Y%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/089583a268ebcac41b5672e45d395703/href">https://medium.com/media/089583a268ebcac41b5672e45d395703/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/802/0*ppm5AZ2dAkKWHQWY.gif" /></figure><p><strong>Note from 2022: </strong>I’ve since updated the receiver and remotes to a “rolling key” scheme to protect against replay attacks, and you should too! The new scheme allows the garage to enroll new physical keys (likely by registering their public key, haven’t looked). I hope to reverse engineer this (and try to break it) in a future post.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e19cf4f0f210" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rotary T9]]></title>
            <link>https://medium.com/@philippp/rotary-t9-eff29ac23868?source=rss-39fe797df35------2</link>
            <guid isPermaLink="false">https://medium.com/p/eff29ac23868</guid>
            <category><![CDATA[hardware]]></category>
            <category><![CDATA[retrocomputing]]></category>
            <category><![CDATA[cyberpunk]]></category>
            <category><![CDATA[arduino]]></category>
            <dc:creator><![CDATA[Latent Cybernetics and Praxis]]></dc:creator>
            <pubDate>Sun, 12 Sep 2021 06:09:30 GMT</pubDate>
            <atom:updated>2021-09-12T19:50:40.957Z</atom:updated>
            <content:encoded><![CDATA[<p>Rotary T9: A detached bridge across the recent past.</p><p>For my 40th Birthday, my brother gifted me a rotary phone dial. Not just any rotary telephone dial — but one that had been in our maternal grandmother’s house since I can remember.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XO5srUkspZHs2iCAq8hpZA.jpeg" /></figure><p>As a visiting high-schooler in the 1990s, I dialed a lot of numbers: An 8-digit prefix that would re-route the call to afford a cheaper international rate, followed by the 11-digit number (country code, area code…). Often I’d have to call back, repeating the process until serendipity put the recipient close enough to their phone. Email or anything internet-related required a 20-minute bike-ride to the local cyber-cafe, but by itself would not have sustained any high school relationships of that era. This made me wonder: <strong>What if I could have texted from Grandma’s rotary phone?</strong></p><p><strong>The hardware interface</strong></p><p>I figured I’d get started on interfacing this device. The design is pretty darn simple and resilient: Four wires come out of the dial:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/866/1*rDxwPe5MAyrRWkgHp-nQEw.png" /><figcaption>White for V-, Green for V+, Yellow is the inverse of “dialer click,” and brown the inverse of “dialer active”.</figcaption></figure><p>To dial a number, the user inserts a finger into the corresponding hole on the face of the dial and rotates the dialer counter-clockwise until the black bar is reached, extending a spring within the dial. When the finger is extracted, the face of the dial rotates back to its original position and “clicks” once for every number that is passed. Counting these clicks reveals the number that was dialed. The contraption on the lower right of the photo above contains the switches controlling the brown and yellow wire’s connectivity.</p><p>The brown wire is connected to the white wire (V-) while the dialer is in motion. The yellow wire connected to the green wire (V+) <em>except</em> when a number is passed during rotation. These disconnections are the “clicks” we count to determine what number was dialed.</p><p>To interface this to an Arduino’s digital IO pins, I connect the brown wire to V+ through an 18kOhm resistor, and the yellow wire to V- through another 18kOhm resistor. I then connect the brown and yellow wires to two arbitrarily chosen Arduino IO ports. When the dialer is at rest, the brown port will be high — and will be pulled low while we are dialing. The yellow wire’s port is also held high, and is pulled low once for every number that is passed.</p><p>I interfaced this with an Arduino Mega with an LCD alphanumeric header, as seen below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5Q9tDL3PNyjoopJrg8KCSg.jpeg" /></figure><p><strong>The Software</strong></p><p>The software can be segmented into low-level input handling (e.g. what number was dialed?), system state (are we dialing numbers or T9 text), and T9 interpretation.</p><p>Low-level input handling requires sampling the brown and yellow wire states at the Arduino’s sample rate. Since the yellow (“click”) pin will be low for multiple Arduino sample periods for even a single “click”, I count the number of transitions from high to low <em>while the active pin is low</em> to determine the dialed number.</p><p>The system state can be changed from “number” to “T9” by using the direction pads, which also changes the active row (the phone number is entered on the top row, the T9 text on the bottom).</p><p>When a new number is received, we check whether we are in number or T9 mode. In number mode, we simply print the number and advance the cursor to the next column (i.e. the next input field). In T9 mode, we define a lookup table (2 is abc, 7 is pqrf, etc) and track how often each number was dialed: If we are dialing “3” for a the first time we print a “d,” if we dial “3” a second time we print an “e,” if we then dial a “4” we advance to the next column and print a “g.” If you need to enter two of the same letter, you must wait for the auto-advance timer.</p><p><strong>The end(?) result</strong></p><p>We now have a dialer that can enter and display numbers or text. I stopped short of adding a Twilio HTTP client and T9 suggest (the best T9 functionality IMHO) and other bells and whistles, since these are reasonably well-solved and arduous, and I feel like this stands on its own as a weekend exercise :)</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F5zz0MWXlSWc&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D5zz0MWXlSWc&amp;image=http%3A%2F%2Fi.ytimg.com%2Fvi%2F5zz0MWXlSWc%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/b60882afc7833fb7c40d50ba23e544c2/href">https://medium.com/media/b60882afc7833fb7c40d50ba23e544c2/href</a></iframe><p>Not cleaned up and unsupported code can be found at <a href="https://github.com/philippp/t9_dialer/blob/main/dialer.ino">https://github.com/philippp/t9_dialer/blob/main/dialer.ino</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eff29ac23868" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Privacy and Cybercrime]]></title>
            <link>https://medium.com/@philippp/privacy-and-cybercrime-c6a81cf6a989?source=rss-39fe797df35------2</link>
            <guid isPermaLink="false">https://medium.com/p/c6a81cf6a989</guid>
            <dc:creator><![CDATA[Latent Cybernetics and Praxis]]></dc:creator>
            <pubDate>Sat, 10 Oct 2020 16:03:08 GMT</pubDate>
            <atom:updated>2020-10-10T18:36:02.696Z</atom:updated>
            <content:encoded><![CDATA[<p><em>These views are mine, independent of those held by my employer.</em></p><p>As a consumers and citizens, we want to maximize privacy.<br>As a defensive security engineers, we need to identify bots.</p><p>Regrettably, the same mechanics — cookies, browser properties, IPs — that allow users to be tracked across sites (violating privacy assumptions of users) are also the easiest way to identify bulk accounts and automated interactions: “bot activity”.</p><p>These forces are on a collision course: Major browsers and hardware vendors are advancing consumer privacy by reducing site operators’ ability to track devices and users. <strong>Without parallel innovations in the detection of scaled abuse, we risk a world in which only the most well-funded platforms can neutralize scaled threats.</strong></p><p>The path of least <em>incremental</em> resistance will be to double-down on alternative and novel ways of fingerprinting: For example, fingerprinting a client by its browser and network properties in lieu of traditional cookies. This will be a war of attrition, as consumer demand will force browsers and operating systems to further occlude or fuzz these properties.</p><p>Defensive teams will spend an increasing share of their time searching for new idiosyncrasies (e.g. how the browser/OS/hardware <a href="https://browserleaks.com/webgl">renders things</a>), and less time chasing cybercrime groups that focus their energy on infra and opsec.</p><p>At this point it is fair to question our reliance on this invasive fingerprinting.</p><h3>Motivating threat: Scaled account abuse</h3><p>Arguably the most common application of client fingerprinting is to detect bulk <em>creation</em> or <em>administration</em> of accounts. Bulk accounts are used for spam, fraud, disinformation campaigns, and other abuse that requires a degree of scale. In these operations, the content created by these accounts may be benign (e.g. “Candidate A for president! Candidate B is corrupt!”); the deceptive and scaled nature of the authoring accounts make it abusive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/1*MEw1zIfJfTCZ6NZzY0YbYw.png" /><figcaption>“Astroturf? No, this real organic activity.”</figcaption></figure><h3>Alternative: Behavioral classification</h3><p>It is possible to build a statistical model of user interactions and sequences and use various unsupervised and semi-supervised methods to identify outlier groups. This may be the most privacy-friendly approach, although it heavily relies on ML expertise and discipline towards auditable enforcement and understanding.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/412/1*HLcbwHpOBex5L4tT-EqAZQ.gif" /><figcaption>Nontransparent enforcement systems are a terrible master.</figcaption></figure><h3>Alternative: Proof of (real-world) existence</h3><p>Another popular approach is to verify real-world artifacts: non-VOIP phone numbers, physical addresses, government IDs, and more.</p><p>In this arrangement, the user is required to trust the service provider enough to provide some aspect of their identity. The service provider benefits by externalizing the cost of verification: Verifying a bank account leverages the bank’s identity verification practices.</p><p>This requires revealing aspects of one’s identity, something enough users take for granted to be a viable product solution. However, citizens of governments that stifle free expression risk physical harm by linking online and real-world identities.</p><p>Another challenge is the non-transparency around data sharing agreements — advertisers may not be able to fingerprint my browser, but the service provider may still sell my identity-linked profile and site history to an aggregator.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/550/1*Otr7cZeSgYQIs0mSWiwzXw.jpeg" /><figcaption>“We adhere to know-your-customer principles”</figcaption></figure><h3>Alternative: Normalize scaled activity</h3><p>The “do nothing” option requires the audience to become accustomed to a maelstrom of sock-puppet accounts and their messages, and create smaller circles of trust. The “prevailing opinion” heard on the high street (e.g. trending page) is expected to be that of the biggest spender, and discriminate consumers of information will need to filter assumptions through diverse and disconnected networks of trust.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*jhtMNYfBXrtV_JesFKn_Vw.jpeg" /><figcaption>YogaDad{1..999} getting ready to upvote and comment</figcaption></figure><h3>Conclusion</h3><p>The future will likely hold a mix of the above, with major platforms leveraging both behavioral analysis and real-world artifacts to manage scaled abuse. Relying solely on identity verification has its pitfalls: Attackers are motivated, and may exchange bottles of vodka for accounts in remote mountain villages sooner than we think.</p><p>The problem will also shift shapes — against the backdrop of a broadening global wealth gap, we are already observing local industries of for-pay online activity, leveraging real devices and humans that intentionally obfuscate their coordination.</p><p>Users thus must become increasingly skeptical, and remember that platforms are motivated to work on these problems to the extent that they are aware of the problem. As citizens we must continue curating our information filtering networks and remain vocal about scaled attempts to manipulate and defraud us.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c6a81cf6a989" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Identifying periods of continued change]]></title>
            <link>https://medium.com/@philippp/identifying-periods-of-continued-change-bcd993e5ea58?source=rss-39fe797df35------2</link>
            <guid isPermaLink="false">https://medium.com/p/bcd993e5ea58</guid>
            <dc:creator><![CDATA[Latent Cybernetics and Praxis]]></dc:creator>
            <pubDate>Thu, 10 Sep 2020 17:05:36 GMT</pubDate>
            <atom:updated>2020-09-10T17:05:36.427Z</atom:updated>
            <content:encoded><![CDATA[<p>I’d like to accurately capture “events” of interest: segments of continuous motion. Naively, I use the localized change detection algorithm and grab a frame whenever I’m done processing the last transition:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FzWAYqsowTn4%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DzWAYqsowTn4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FzWAYqsowTn4%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/63f2a801c64b2e280d7a5af3429fdc37/href">https://medium.com/media/63f2a801c64b2e280d7a5af3429fdc37/href</a></iframe><p>Not too terrible for zero effort! Let’s try it with a person (c’est moi) staggering in front of the camera:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FPbVNi9yihzU%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DPbVNi9yihzU&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FPbVNi9yihzU%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/81a812eca99a674fee200e222a8ffc40/href">https://medium.com/media/81a812eca99a674fee200e222a8ffc40/href</a></iframe><p>…this one could be improved. In this iteration, I’d like to complete the following two objectives:</p><ol><li>Optimize the capture of the bounding box on a per-frame basis</li><li>Identify and export a motion sequence.</li></ol><p>Multi-object tracking is currently a non-goal, but would be nice in the future — followed by object classification.</p><h4>Bounding Box Optimization</h4><p>We previously used the ratio of the mean RGB-RSS difference within the rectangle to the global mean RGB-RSS difference as a scaling factor for the frame’s variance. We have two variables for this optimization: (1) the scaling factor — how far we downsample the image for box detection and (2) interest threshold — how many standard deviations above the norm does an RGB-RSS pixel need to be in order for us to “grow” the box.</p><p>It should be noted that our video feed is more “robust” than the 30 second screenshots: Fewer ambient pixels change in each sub-second capture, but we also miss the action during several transitions:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M1QZY5yvMZQyJ5Auj4Lbqg.png" /></figure><p>As a first attempt, I’ll try dynamically adjusting the RGB-RSS threshold to maximize the rectangle’s heat capture.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bcd993e5ea58" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>