A behavioral analysis company needed a high-quality video dataset to train face and voice sensing AI for mental health and emotional-state assessment. The effort required sustained, recurring data collection rather than a one-time capture. The dataset also needed to reflect diverse participant characteristics across facial shapes, sizes, skin tones, and ages. A monthly video collection program was implemented to gather participant recordings for AI training. Participants recorded short videos that captured both facial and voice signals for analysis. The collection process continued over roughly a year to build a consistent dataset for model development. The program resulted in a dataset of more than 1,000 participant videos for training facial recognition and voice analysis algorithms. Videos were collected on a monthly cadence, with each recording lasting a few minutes. The final dataset included adults across varied demographics to better support broad clinical and patient use cases.

Skills

Project Details

A security technology company needed better training data to improve its threat detection algorithms. Existing datasets did not provide enough realistic footage of violent incidents from security camera viewpoints. The client required consistent perspectives and technical specifications to support model training and downstream labeling. Custom video datasets were produced featuring actors performing violent scenes. The production created 25 videos per group, each 30 seconds to 1.5 minutes long, depicting fights, shoving, and strangling from security camera perspectives. Videos were delivered in 720p at 24fps with static camera angles positioned 4–5 meters high. The client received standardized video assets tailored for threat detection training. Delivery specifications ensured consistent quality and camera perspective across the dataset. The client began the annotation process immediately after receiving the videos.

Skills

Project Details

Researchers at a US-based research institute needed high-quality video data to train AI to label human actions at mass scale. They also needed data to train long-range biometric identification. Existing sources such as YouTube had often produced low-quality or not fit-for-purpose videos for this type of research. The institute commissioned a large freelancer network, referred to as the “human cloud,” to record videos of everyday activities such as getting into cars and exiting buildings. The project recruited participants at scale and coordinated collection of consistent video data for research needs. A secure vault payment system was used to manage international payments to freelancers. The collected data enabled the researchers to create an open-source video library for the international research community. The resulting library provided an alternative to relying primarily on YouTube for similar datasets. The project provided an estimated 20,000 freelancers with opportunities to participate in paid work during the COVID-19 outbreak.

Skills

Project Details

HyperSentience needed authentic audio recordings of everyday activities captured in real home environments. The goal was to produce diverse, naturally-captured audio data to train context-aware AI models. These models were intended to help devices like smartphones, VR/AR headsets, and smartwatches understand user activities and adapt accordingly. The team also needed the dataset to reflect real-world variability in capture conditions. A data collection effort was implemented to gather authentic audio recordings of individuals performing everyday activities at home. The recordings were captured to support training context-aware AI models for activity understanding across consumer devices. The dataset was organized to include multiple capture variables such as phone distance, mounting positions, location specifics, and background noise types. The project also included step-by-step documentation to support efficient annotation. The effort delivered naturally-captured audio data intended for training activity-aware AI models. The dataset included explicit distance groupings and other contextual variables to improve representativeness. It also provided documentation designed to streamline downstream annotation work. Secure payment processing was handled through a vault system.

Skills

Project Details

HyperSentience needed authentic audio recordings of people performing everyday activities in real home environments. They aimed to generate diverse, naturally captured audio data suitable for training context-aware AI models. The effort required capturing realistic variations that typical lab-style recordings could not provide. A data-collection project was implemented to gather in-home audio recordings of individuals completing daily activities. The dataset design accounted for multiple recording variables such as device distance, mounting positions, location specifics, and background noise types. Documentation was produced to support efficient annotation of the collected data. The initiative delivered a comprehensive dataset of naturally captured home-environment audio recordings for model training. It captured multiple environmental and device-placement variables to support context-aware AI development for devices like smartphones, VR/AR headsets, and smartwatches. It also included defined distance groupings and accompanying documentation to streamline downstream labeling.

Skills

Project Details

A security technology company needed realistic training footage to improve its threat detection algorithms. Existing datasets did not provide enough controlled scenarios from true security-camera perspectives. The company required consistent clips depicting violent interactions for model training. Custom video datasets were produced using actors performing staged violent scenes. Each group included 25 videos, with clips ranging from 30 seconds to 1.5 minutes and depicting fights, shoving, and strangling. Footage was recorded from static security-camera angles positioned 4–5 meters high. Videos were delivered in 720p at 24fps. The company received grouped datasets that matched the required scenarios and camera constraints. Delivered clips provided consistent security-camera-style perspectives and standardized technical specifications. The customer began the annotation process immediately after delivery.

Skills

Project Details