The Task
A company developing a real-time video analytics system for public safety approached us with an urgent task: to create a custom dataset of realistic fight scenes filmed in various public and indoor environments. No existing open-source datasets met their criteria — everything had to be built from scratch.
The brief was ambitious: 200 high-resolution videos, filmed in diverse settings with natural-looking scenarios, and delivered within two weeks. With our strong track record, the client trusted Unidata to make it happen.
The Solution
-
- 01
-
Planning and Pre-Production
We began by designing a detailed production strategy. Each video had to be filmed in 4K with wide-angle coverage and include multiple participants to simulate real-world conditions.
Our team developed scene scripts and selected varied indoor and outdoor locations to ensure data diversity.
-
- 02
-
Casting and Coordination
We partnered with local boxing clubs in the city to cast trained fighters capable of safely executing realistic fight choreography.
Filming took place in controlled environments, with attention to safety, scene dynamics, and visual clarity.
-
- 03
-
Weathering the Unexpected
An unplanned snowstorm during outdoor shooting presented serious challenges — from rapidly draining camera batteries to reduced actor endurance. Our team responded quickly, using power banks to stabilize equipment and keeping the cast warm and ready.
While ten scenes required reshooting, the team stayed agile and kept the project on track.
-
- 04
-
Quality Control and Dataset Consistency
To ensure the dataset met the client’s technical standards, we used multiple camera setups for each scene, capturing every event from several angles.
We implemented regular quality checks throughout the shoot, maintaining consistency across all materials.
The Result
On-Schedule Delivery: All 200 videos were delivered within the two-week timeframe.
Data Quality: The dataset featured high-resolution, multi-angle footage suitable for training sophisticated machine learning models.