When Training AI Means Throwing Punches: 3DiVi Engineers Stage Fights to Teach Violence Detection

To make Artificial Intelligence (AI) smarter—and streets safer developers at U.S.-based computer vision company 3DiVi have taken an unusually hands-on approach: simulating real physical fights themselves to train AI how to recognise violence.

Known for its face recognition and body tracking technologies, 3DiVi faced a critical roadblock in developing a new AI model for detecting aggressive behaviour: a lack of real, annotated fight footage.

To create a production-ready system, the team needed a dataset covering a wider variety of scenarios, gestures, and escalation patterns. But sourcing that kind of footage was not easy—not because it does not exist, but because legal, ethical, and practical constraints limit both access and variability of real-world samples.

“We had solid fight data, but it wasn’t enough to confidently validate our models across all scenarios,” said Leonid Leshukov, Head of Product Development at 3DiVi. “So, we did the next best thing: staged our own brawls.”

With cameras rolling from multiple angles, team members choreographed mock street fights, mimicking aggression, stumbles, and crowd reactions. The footage is now being used to train AI capable of detecting violence and threats in public places—a crucial component in modern surveillance and smart city systems.

The approach also addresses a known flaw in many existing datasets: they are too staged, too clean, or simply not reflective of real-world conditions. By creating their own, 3DiVi gained full control over the context, accuracy, and diversity of the data—ensuring the AI learns from something closer to reality.

“It’s not just about seeing a punch,” Leshukov added. “It’s about teaching the system to notice movement patterns—body posture, escalation, proximity—and respond in time.”

The technology is already being piloted with video surveillance platforms and smart city infrastructures.

And no, in case you are wondering—no developers were harmed in the making of this dataset—just bruised egos and a lot of laughing.