UNDO CCTV detection

UNDO CCTV detection algorithm

For the past year the project has been planning and implementing workshops with activist communities in Malmö, Helsinki and Copenhagen.

The aims of these workshops are multiple. First and foremost, we try to understand and work around the restricitons set up by institutions, namely the police, in researching ubiquitous surveillance of urban spaces. Then, we try to gather requirements for potential software solutions that:

1. Could be of use to activists communities

2. Could help with research and understanding the state of the surveillance infrastructure deployed

So, moving towards this direction we have trained and developed an algorithm that can detect CCTV in provided images. This is a small fisrt step and we plan to use this simple detector on a bigger system that can detect, analyse and provide detailed descriptions of the surveillance cameras.

We trained a YOLO8 model on our custom dataset for 20 epochs. Most images used in training have been collected by us through ethnographic research, some images used in this project come from the dataset of the Fuziih CCTV-Exposure.

The model so far has the following detection metrics:

Precision and recall have reached balanced high scores (above 83%).
mAP@0.5 (object detection quality at 0.5 IoU threshold) reached 87%.
mAP@0.5–0.95 (stricter localization accuracy) reached 41%.

results

The code for the detector is open source and lives over at GitHub:

https://github.com/jethronap/cctv-detection-UNDO

A preview of the application that can be tested is hosted at HuggingFace:

https://huggingface.co/spaces/jnap/UNDO-project

Please contact us if you want the dataset or the weights of the model.

The UNDO team.