New Delhi: Having maintained that over 2,800 people with “criminal antecedents” were part of the Cockroach Janta Party (CJP) protest in July and were “identified” using a facial recognition system (FRS), Delhi Police ’s own officers say the accuracy of the system is between 60% and 80%, and if an image that is captured is unclear, it can return matches with similarity scores as low as 16% to 20%.
However, they did not respond to allegations that some people flagged by the system were in jail on July 20, when the Sansad Chalo march was called by CJP. Delhi Police has a database of more than 2.5 lakh people from across the country. Officers set varying accuracy thresholds, such as 80% or 60%, but an FRS match serves strictly as an investigative lead rather than conclusive court evidence, police said. The officers said the system can generate up to 12 potential suspects or matches for a single image. In some cases, particularly when the source image is unclear or captured under difficult conditions, the system can return matches with similarity scores as low as 16% to 20%.
Delhi Police officials Friday said they had not arrested any individual with criminal antecedents based on the FRS report and that an exercise was under way by local police stations to verify the whereabouts of the people flagged by the system. The presence of each individual, police said, will be corroborated with prison details as well as mobile data locations, CCTV footage and witness testimony. Some may even be out on bail and still wrongly flagged, they said. Officials said the technology is particularly useful in crowded areas and during major deployments, where manually identifying individuals from hundreds of people can be difficult. It has also played a major role in tracing missing people and children, police said.
The system analyses several facial parameters and assigns a similarity score to the possible matches.
The software then analyses the facial features and generates potential matches, which can provide investigators with leads about the identity of a suspect,” officers said. “When investigators obtain a photograph or a clear frame from video footage, the image is fed into the system and compared with faces stored in the database.