header object
Ask Expert
header bars header close
Computer Vision down arrow down arrow active
Services down arrow down arrow active
Image annotation Video annotation 3D point cloud Automatic annotation Data validation Semantic segmentation Data creation Data collection Generative AI Custom AI Data compliance Physical AI & robotics data
Industries down arrow down arrow active
Automotive Aerial Agriculture Construction Documents Fashion Livestock Logistics Manufacturing Maritime Medical Retail Robotics Security Sport Waste management
LLM down arrow down arrow active
Services down arrow down arrow active
Data for agentic AI models Virtual environments & RL-gyms Data for creative AI Data for reasoning skills AI safety & red teaming Solutions for coding agents VLM & multimodal AI solutions LLM agents training
Industries down arrow down arrow active
Automotive Agriculture Business Education Finance Gaming Human resources Marketing Medical Real estate Retail Robotics Security Sport Sofrware development
Resources down arrow down arrow active
About us Keylabs platform Case studies Conferences F.A.Q. Media Podcasts Blog
fb
linkedin
youtube
instagram
x
SeeChange retail theft detection case study header

How Keymakr supported
SeeChange in training AI
systems for retail theft
detection

rsip vision cipia v7 AMD shopify walmart logitech Recycleye viisights jabra zkt qm qm qm qm qm qm qm

AI-powered solutions for retail

Download PDF
Company:
Services:
Video annotation icon Video annotation icon hover
Video annotation (bounding boxes)
Event-based labeling icon Event-based labeling icon hover
Event-based labeling
Dataset filtering icon Dataset filtering icon hover
Dataset filtering
Overview:
2 months case study period
Up to 50 specialists and 2.5M video frames processed
Up to 50 specialists and 2.5M video frames processed

Intro

SeeChange develops AI-powered solutions that help retailers detect and prevent theft, improve operational visibility, and reduce losses in physical stores. By applying computer vision and machine learning to existing in-store camera systems, the SeeChange platform transforms traditional surveillance into intelligent retail analytics that can identify suspicious behavior and monitor product interactions in real time.

The company's technology focuses on understanding how customers interact with products throughout the shopping journey. Its AI models can recognize events such as items being taken from shelves, placed into baskets or bags, carried across the store, or handled in ways that may indicate potential loss scenarios. As a result, retailers can better understand store activity and respond quickly to emerging risks.

The platform operates in complex, real-world retail environments, where large volumes of video data must be processed and accurately interpreted. To achieve reliable detection and behavioral analysis, SeeChange relies on high-quality training datasets that capture real customer interactions with products and shopping objects across diverse store layouts and scenarios.

The collaboration with Keymakr focused on preparing such structured datasets that enable AI systems to learn to detect key interaction events.

Retail store environment with customer-product interactions for AI training

The challenge

SeeChange provided a large volume of video data from in-store surveillance cameras, capturing a wide range of customer behavior scenarios within retail environments.

The main challenge was that most video frames lacked relevant events. Customers could simply walk through aisles or browse shelves without interacting with products.

The initial dataset contained approximately 2.5 million frames, so manually identifying and annotating relevant events required a well-organized workflow.

The key objectives of the project were to:

  • identify relevant product-interaction events within the video data
  • annotate the objects involved in these interactions
  • prepare structured datasets for training behavioral analysis models.
Challenges of retail video annotation with large datasets and sparse relevant events
The solution

To efficiently process the large volume of video data, the Keymakr team designed a multi-stage workflow that combined dataset filtering, event-based annotation, and quality control.

Stage 1. Dataset filtering

The first step involved a complete review of the raw video dataset, which contained approximately 2.5 million frames captured by surveillance cameras in retail stores.

The Keymakr team carefully reviewed the entire dataset to identify frames where customer-product interactions actually occurred.

This approach made it possible to:

  • significantly reduce the volume of data for further annotation
  • focus only on frames relevant for training behavioral analysis models
  • optimize the time required for labeling

As a result, the dataset prepared for annotation was reduced by approximately 5 times, substantially accelerating the entire workflow.

Filtered dataset with relevant customer-product interaction frames
Stage 2. Event-based labeling

Using an event-based approach and identifying key moments of interaction between customers and products, the team annotated the objects involved in these interactions.

The dataset followed a defined object taxonomy used to train the product-detection model. In total, 12 classes of retail interaction objects were identified.

The labeling schema also differentiated between relevant product-related interactions and other objects, helping improve model accuracy and reduce false detections.

Bounding boxes were used to annotate objects, enabling precise capture of relevant interaction moments. Annotations were applied selectively, ensuring that only meaningful instances were included in the dataset.

To maintain consistency and quality, standardized annotation guidelines were drafted and followed, helping ensure reliable and accurate training data.

The project was carried out using Keylabs, Keymakr's proprietary annotation platform. To process the large volume of data within the required timeframe, Keymakr assembled a core team of 30 annotators that scaled up to 50 at various stages of the project.

Event-based labeling workflow with standardized annotation guidelines
Handling complex scenarios

In most cases, product interactions were clear enough to annotate. However, some complex scenarios occasionally arose due to the characteristics of the video footage.

In these cases, the team applied contextual analysis across frame sequences to accurately identify interaction moments, ensuring that relevant events were captured even when individual frames provided limited information.

This approach helped maintain both consistency and reliability across the dataset, supporting the creation of high-quality training data for behavioral analysis models.

The annotated datasets were delivered in a structured format compatible with the SeeChange's machine-learning pipelines, enabling seamless integration and efficient use in model development and iteration.

Timeline

Since retail theft directly impacts store revenue, SeeChange's often require rapid deployment of AI solutions.

As a result, the project was carried out within very tight timeframes of 2 months, sometimes with deadlines that required maximum team mobilization. The Keymakr team did everything possible to deliver results as quickly as possible, while maintaining the required level of accuracy and attention to detail.

Results

Thanks to a structured annotation workflow and the Keymakr team's scalability, high-quality training data was prepared for SeeChange's AI systems that analyze customer behavior and help retailers detect potential theft scenarios.

Optimized dataset for model training
A preliminary review of the video data reduced the original dataset by approximately 5 times, focusing only on frames containing relevant product-interaction events. This significantly accelerated the preparation of data for training computer vision models.
2.5 million frames analyzed
The Keymakr team reviewed approximately 2.5 million video frames from surveillance cameras to identify events relevant for training theft-detection systems.
100,000 objects annotated
During the project, over 100,000 objects were labeled for the dataset, including products, baskets, bags, and other items that customers interact with.
Fast delivery under tight deadlines
Since SeeChange's solutions help retailers reduce theft-related losses, data preparation often requires rapid turnaround. The Keymakr team assembled 50 annotators and ensured the timely delivery of results within strict deadlines.
"Keymakr has proved to be a very reliable partner in our process for creating our machine learning products. They have quickly become proficient in annotating complex images and have been very responsive and quick to complete our projects. Thanks for your support, Keymakr!"
Quote icon

Hoomi Chadirchi, SeeChange Chief Architect

"I really appreciate projects like this. SeeChange focuses on fundamental, real-world needs and consistently delivers practical, meaningful solutions. This approach is especially important, as it demonstrates how AI creates value. In fact, these kinds of projects play a critical role in making AI more accessible, understandable, and impactful, effectively solving both everyday challenges and more complex problems at scale".
Quote icon

Zoya Boyko, Keymakr PM

Reviews
on

down-line
g2
star
star
star
star
star

"Delivering Quality and Excellence"

The upside of working with Keymakr is their strategy to annotations. You are given a sample of work to correct before they begin on the big batches. This saves all parties time and...

Read more
star
star
star
star
star

"Great service, fair price"

Ability to accommodate different and not consistent workflows.
Ability to scale up as well as scale down.
All the data was in the custom format that...

Read more
star
star
star
star
star

"Awesome Labeling for ML"

I have worked with Keymakr for about 2 years on several segmentation tasks.
They always provide excellent edge alignment, consistency, and speed...

Read more

More
cases

down-line
Confidential client icon

Automotive

Delivering scalable traffic detection data for a major automotive company

Read more
SoundAware logo

MusicTech

Providing consensus-based media monitoring to help SoundAware protect the intellectual property rights of creators

Read more
Cognex logo

Robotics

Helping Cognex with high-quality OCR and object detection datasets for industrial automation

Read more