In the Parisian demonstrations against the pension reform, the three sources systematically announce different figures. The question of the assessment of crowds in processions is an endless debate.
How do you count a crowd demonstrating? And especially who to believe? During the Paris demonstration on Friday, the police, for example, counted 31 000 people, the CGT between 350 and 400 000. These deviations, repeated, sometimes of the order of 1 to 10, or even more, between the police and the organizers, convinced part of the media to appeal, since the end 2017, to another source. “It was not journalistically satisfactory to give two alternative truths,” explains Thomas Legrand, political editorialist at France Inter, behind, with others, this pooling of the figure.
Close to 80 media pay to benefit from a figure supposed to escape any suspicion of manipulation and delivered by a research firm based in Paris. “The only figure that is not political and that comes closest to the truth,” assumes Thomas Legrand. For the Friday event, Occurrence counted for example 39 000 people. On average, out of thirty events counted by Occurrence, their figures differ by a little more than 15% of those of the police, more or less, and close to 250% of those of the organizers, always below.
While the police and unions count manually, by clicking on different parade points, Occurrence uses automatic detection. Placed in a hotel room high up on the course of the demonstration, a camera films the demonstrators' passages. The flow is then analyzed by an algorithm which counts people crossing a virtual line directly above the camera, all cut into slices of 10 seconds.
To compensate for the technical failures of the algorithm, which bug when the passages are very compact, Occurrence performs adjustments as a function of density, and recounts certain passages – the densest – by hand. On average, the manifestations are the subject of 20% recovery. But now, in conclusions published mid-December 2019 on Twitter, Bruno Andreotti, professor of physics at the University of Paris VII, invited in May 2018 to observe an Occurrence count, was not kind. Measurement problems, too significant recovery, disregard of demonstrators escaping the camera …
“The raw measurement that comes out is not good. It breaks its mouth when it's very dense, the algorithm doesn't work. It's like having a thermometer stuck at 45 degrees and you wanted to take your chicken oven temperature at 120 degrees, “he said. A salvo little appreciated by Assaël Adary, the boss of Occurrence, who considers these critics “unfair”. Occurrence does not take into account the demonstrators joining the procession without going through their counting point, those arriving by adjacent streets or who join the demonstration at the end. “We count the people who pass our sensors, including bystanders on the sidewalks who do not demonstrate. But not the others, “he explains.
” This raises questions “
” The algorithm is imperfect because its reliability will depend on the density of the crowd. And if we couple with a recount by hand, it is problematic “, estimates for his part Medhi Moussaid, researcher specialized in crowd behavior. CGT side, the arrival of this new actor, who gives figures always lower than his, is not necessarily appreciated.
“I find their figures incredibly low, lower than those of the police. This raises questions, “said Benoît Martin of UD CGT Paris. “I find that their displayed impartiality does not correspond to reality”. Despite what he calls “flaws”, Mehdi Moussaid assures that “the general principle is good”. “We had to do with wild counts, with cognitive biases, uninformative. The principle of more neutrality is good. And, like every technology, it evolves, version 1 is bad, but version 2 will be better. ”