🧵: Avoiding exposure to facial recognition technology in NYC is near impossible, our research published today shows.
We enrolled nearly 6K digital volunteers to map surveillance cameras at intersections across New York City!
How did we do it?
amnesty-crisis-evidence-lab.github.io/decode-surveil…
We enrolled nearly 6K digital volunteers to map surveillance cameras at intersections across New York City!
How did we do it?
amnesty-crisis-evidence-lab.github.io/decode-surveil…
Volunteers participated via @Amnesty Decoders, which uses microtasking to help researchers answer large-scale questions.
Volunteers were shown a Google Street View panorama of an intersection and asked to find surveillance cameras… 📸
decoders.amnesty.org/projects/decod…
Volunteers were shown a Google Street View panorama of an intersection and asked to find surveillance cameras… 📸
decoders.amnesty.org/projects/decod…
We asked volunteers to record what cameras were attached to. They were presented with three options:
1⃣street light, traffic signal or pole
2⃣building
3⃣something else
If they selected 1⃣, they were asked to identify the camera type.
🎨 Illustrations by Eliana Rodgers
1⃣street light, traffic signal or pole
2⃣building
3⃣something else
If they selected 1⃣, they were asked to identify the camera type.
🎨 Illustrations by Eliana Rodgers
A tutorial video, visual help guide, and moderated forum ensured that volunteers were supported throughout.
On average each volunteer took 1.5 minutes to analyze an intersection. ⏲️
📽️ Video produced by Fabio Basone and presented by Rebecca Echevarria
On average each volunteer took 1.5 minutes to analyze an intersection. ⏲️
📽️ Video produced by Fabio Basone and presented by Rebecca Echevarria
Info about camera placement + type was used as proxy for public or private ownership
For the research it was assumed that dome cameras attached to street lights, traffic signals or roadside poles, were likely to be owned by a gov agency w/ the permission + access to install them
For the research it was assumed that dome cameras attached to street lights, traffic signals or roadside poles, were likely to be owned by a gov agency w/ the permission + access to install them
Each intersection was evaluated by 3 volunteers 👪
The data we published today is the median camera count per intersection. We chose the median as it's least susceptible to outliers + individual errors!
Counts were normalized by the no. of intersections in each neighbourhood…
The data we published today is the median camera count per intersection. We chose the median as it's least susceptible to outliers + individual errors!
Counts were normalized by the no. of intersections in each neighbourhood…
We also used a qualitative review by an "expert" (cue @SofyCaf!) to look for systematic errors or biases, such as vols tagging only one of two cameras on NYPD Argus boxes. 🙈
This preliminary analysis ensured correct orders of magnitude, sufficient for early conclusions! Phew.
This preliminary analysis ensured correct orders of magnitude, sufficient for early conclusions! Phew.
Publishing while the project was live required drafting in extra support and expertise. The data sci team, @JCornebise and @MatchaPillai, take full cred for the analysis above!
The maps were made by @davidcblood 🗺️
We also worked w/ @mjmarci to build a 3D site model…
The maps were made by @davidcblood 🗺️
We also worked w/ @mjmarci to build a 3D site model…
3D modelling was used to demonstrate the distance from which an NYPD Argus camera might be able to capture video processable by facial recognition software.
In the absence of tech specs for the NYPD Argus cameras, we researched commercial models…
In the absence of tech specs for the NYPD Argus cameras, we researched commercial models…
Historical Google Street View imagery told us that the camera we modelled was installed between Nov 2017–Jun 2018
Data on commercial cams available at the time suggest that it was likely to be a Pan, Tilt, Zoom model containing a 6–134 mm varifocal lens able to film at 4K or 8MP
Data on commercial cams available at the time suggest that it was likely to be a Pan, Tilt, Zoom model containing a 6–134 mm varifocal lens able to film at 4K or 8MP
The 3D model shows that the camera could potentially monitor neighbouring roads and capture faces in high definition from as little as few metres to up to 200 metres (or two blocks) away!
The entire project was underpinned by a dataset of traffic intersections generated by @AI_Micah
We were unable to find a complete dataset, so had to create our own. Micah extracted @openstreetmap vector data then processed it using GIS software. Done.
Now let's #BanTheScan!
We were unable to find a complete dataset, so had to create our own. Micah extracted @openstreetmap vector data then processed it using GIS software. Done.
Now let's #BanTheScan!
Actually, not done!
We are yet to evaluate 3 "expert" datasets generated by @vickaytee (who scoped the project!) & 2 vols. In lieu of a "ground truth" these expert datasets will tell us more about the accuracy of the crowdsourced data
More nuanced analysis coming soon.
We are yet to evaluate 3 "expert" datasets generated by @vickaytee (who scoped the project!) & 2 vols. In lieu of a "ground truth" these expert datasets will tell us more about the accuracy of the crowdsourced data
More nuanced analysis coming soon.
In the background but critical to the success of this and past decoders was the careful and thoughtful work of UX/UI designer @diggyd.
Thanks to Dave we user tested again and again and 🔃 ⤵️⤴️🔄⤴️🔃
The designs were developed and put into code by the team @xpondigital 🙏🏾
Thanks to Dave we user tested again and again and 🔃 ⤵️⤴️🔄⤴️🔃
The designs were developed and put into code by the team @xpondigital 🙏🏾
• • •
Missing some Tweet in this thread? You can try to
force a refresh
