Product Counter for Estimating Sales to Assess Credit Worthiness

Demonstration of object detection, tracking and counting

For MSMEs looking to purchase from wholesalers to restock their inventory, access to credit is critical, but can often be limited by the lack of a digitized sales ledger. To compound this pain, smaller merchants see digital point-of-sale systems either as overkill for their shop, or as requiring too much training. Additionally, shop owners with multiple shops have issues with employees miscounting inventory, in some cases intentionally defrauding the shop and stealing goods.

By positioning an Android device with a camera such that it has a view of the counter, we can utilize Tensorflow's computer vision libraries to perform object detection and tracking, counting and tracking the time distribution of how many items pass over the counter in a given day. As the variance in pricing is typically quite tight, an average or median price in the shop can be used to estimate the gross revenue of products sold. Reporting interfaces can be exposed to both the shop's owner, for the sake of reconciliation, and to potential lenders, for the sake of providing data to be used in an assessment of creditworthiness.

Use Case for MSMEs: Product Counter for Estimating Sales to Assess Credit Worthiness

Needs Identified. For MSMEs looking to purchase from wholesalers to restock their inventory, access to credit is critical, but can be limited by the lack of a digitized sales ledger. To compound this pain, smaller merchants see digital point-of-sale systems either as overkill for their shop, or as requiring too much training or effort to operate. Additionally, shop owners with multiple shops have issues with employees miscounting inventory, in some cases intentionally defrauding the shop and stealing goods.

Technology. By positioning an Android device with a camera such that it has a view of the counter, we can utilize Tensorflow's computer vision libraries to perform object detection and tracking, counting and tracking the time distribution of how many items pass over the counter in a given day. As the variance in pricing is typically quite tight, an average or median price in the shop can be used to estimate the gross revenue of products sold. Reporting interfaces can be exposed to both the shop's owner, for the sake of reconciliation, and to potential lenders, for the sake of providing data to be used in an assessment of creditworthiness.

Validation. After visiting dukas in the Kibera district of Nairobi Kenya to demonstrate a proof of concept of the object tracking, we had strong validation of the hypothesized value propositions. While access to credit seemed valuable to each shop owner, the owners of multiple shops saw the reconciliation and monitoring of theft as the technology's primary potential value.

Gap. The main gaps identified were around localized product data and quality of hardware. While the proof of concept generally worked to identify products even in the low-lighting conditions, in order to effectively detect only products in inventory that are passing across the counter, the model should ideally be trained on labeled images of the local products being sold. As far as hardware, while shops do often have smartphones, they are often several generations old, and sometimes in a damaged state. To conduct our testing, we purchased a range of common smartphones that had a camera on them, which would likely need to be offered to merchants on a payment plan.

Potential Partners. To date, we have currently built the demo apps in-house, but are working to organize discussions with the team behind the TensorFlow Android libraries to see what support may be available in including this work as a demonstration in their core libraries. We have also engaged with local developers who have been experimenting in this area, to explore whether they would have interest in taking this product forward. Community banks have also expressed interest in creating credit models based exclusively on the average quantity of items being sold daily.

MSMESCarmen Merab