The Californian technology company builds Artificial Intelligence and Machine Learning models for global SMBs and Fortune 500 companies operating in the retail space. The AI/ML models help in product categorization, visual search, virtual fitting rooms, and provide personalized recommendations to customers.
To build accurate AI/ML models, the company extracts meaningful information from visual data repositories available with their retail partners. They label images to train, validate and build computer-vision based algorithms to recognize objects similar in a way that a human can.
To create an intelligent AI model for a home décor, furniture and fashion client, the company received a training data repository of 1.2 million images. They had to:
- annotate the images using appropriate terms to identify and classify the various products
- prepare an accurate data set with 2D bounding box images
Due to the variety and volume of the data, annotating images for different specifications was slowing their speed to market. Managing a vast data annotation team demanded huge time, training and cost investment.
Faced with a tight deadline and critical need for accuracy, the company partnered with Hitech for tagging of data and consistent flow of labelled dataset to train computer vision model.
- Categorizing and comprehending different types of home décor, furniture, fashion apparels, and accessories from the compiled and raw mix of images
- Managing an input volume of 1.2 million images within a tight deadline of 12 days with quality checks
- Training a skilled and experienced team for the project in a short time
Hitech data specialists classified and annotated more than a million home décor, furniture, fashion apparel and accessory images. Bounding boxes were used to enhance the accuracy of object detection and add identifiers for the model to learn autonomously. The annotated images prepared a perfect training baseline for the client’s AI-model.
- Hired and trained an experienced team of data annotators and quality control professionals with specific industry/domain knowledge.
- SOPs with clearly defined KPIs and other project metrics were defined and documented.
- Working on the client portal through secure credentials, the team first assigned unique IDs to each product image.
- ‘Bounding Box’ tool was used to create rectangular boxes around the images to accurately identify objects.
- Annotated each product within the bounding boxes with relevant terminologies based on pre-defined industry- based tags and keywords.
- Added additional classification and keywords to each home décor, furniture and fashion image during annotation to simplify future identification and retrieval.
- Applied stringent quality check process to verify image batches for expected quality benchmarks; updated the approved batches into the client’s CRM system.
- Completed image annotation and tagging for 1.2 million images within a 12-day timeline.
Automated workflows saved time on data preparation and boosted annotation productivity by 96%
High quality annotated data increased algorithmic accuracy
Annotated training data delivered at scale enhanced model performance and relevance