Computer Vision and Image Recognition
Vispera Recognition Engines are based on state-of-the-art Convolutional Neural Networks and Recurrent Neural Networks
Our engines are specifically tailored for retail scenes, for packaged products and their taxonomy, whereas off-the-shelf image recognition frameworks and APIs lack any domain specific adaptation. This makes our Image Recognition (IR) technology superior to its counterparts in terms of accuracy, repeatability and ability to handle diverse use cases and custom visual tasks.
Vispera Recognition Engines are reinforced by the most sophisticated panorama reconstruction and image segmentation algorithms for retail scene parsing, which maintains our image rejection rates at a record-low, well below 1% over all submitted images.
Vispera Recognition Engines have high tolerance to adverse picture-taking conditions such as non-frontal views of shelves taken in narrow aisles or pictures with severe glass reflections from coolers with doors closed.
Thanks to the hyperscale computing architecture of our cloud IR framework, we can achieve unmatched reporting speeds. Images taken in the field with mobile photoshoot are directly synced from client to cloud, where they undergo several image processing/recognition steps including perspective normalization, scene panorama construction, item detection/recognition, price reading, and KPI computation. Scaled and executed fully automatically, all these steps will take under 5 seconds per image, and including data upload/download overheads, complete reports for typical visits and image sets can be delivered to both mobile and web dashboard under a minute. We can also enable on-device (offline) recognition and reporting, if there is no or limited mobile connectivity in the field. In that case, depending on mobile processing power, stock availability can be reported in less than a second it is executed entirely on-device and without any accuracy compromises.
The baseline performance of Vispera Recognition Engines starts from 93% accuracy for SKU level product recognition, and exceeds above 96% depending on the product category as more training data accumulates for continued machine learning. Engines are further powered by our proprietary human-in-the-loop approach that is actively driven by system recognition confidences. This process not only enables automatic discovery and system enrollment of novel SKUs or SKU packagings as they appear on the shelf, but also delivers above 99% SKU level recognition accuracy with smart and persistent quality control, making us the only IR solution provider in the retail domain who can commit to near-perfect performance levels.
Reporting and Analytics
- Data drilldown powered by ElasticSearch, supporting complex queries with super fast query times
- Highly customizable tagging architecture for clients' master data requirements
- Value added KPIs from raw data through flexible KPI engine
- The most flexible planogram compliance checking solution in the market: Reference-based and rule-based controls
- Planogram and Realogram inspector
- Visualization of data supported by Highcharts
- Easy integration with client systems and 3rd party analytics and prediction services through REST API
- PowerBI framework support for diverse report customization and exports
- Highly flexible, customizable data collection tools
- Mobile app: Android OS & iOS
- Versatile in-store image and data collection process, highly customizable to diverse retail execution and auditing projects
- Hybrid system for flexible data collection, supports manual input when constrained scenes or permission issues interfere with photoshoot
- Online data synchronization, transparent data upload and offline working capability
- Typical shelf complete less than 10 seconds with successive photoshoot feature
- Fast navigation
- Fast photoshoot
- Instant image quality feedback
- Intuitive UI and navigation for correct data collection
- Picture grouping into custom display types
- Guided photoshoot
- Reduced human errors with built-in checks
- Augmented reality features for on-device panorama reconstruction
- On-Device product recognition and OOS detection
- User authentication
- Date and time logs
Mobile App Integration
- Cloud solution and recognition engines backed by Azure
- Scaling by Elastic Beanstalk
- Secure multi-tenant and dedicated private instance options
- GPU and CPU based inference
- On site infrastructure solutions thanks to our partnership with partnerships with Intel, Nvidia, HP Inc, Cisco, Dell, Microsoft