Counting Shelf Facings with 96.8% Accuracy in Hong Kong

A multimodal retail execution program measured dense shelf facings, complex weekly promotions, and new-product speed to shelf across major chains and independent pharmacies.

Overview

Measure dense shelves, layered promotions, and new-product rollout in one workflow.

A global consumer goods company needed reliable execution data across Hong Kong’s major supermarket and personal-care chains as well as hundreds of independent pharmacies. Clobotics combined fine-grained facing recognition, promotion-text interpretation, and speed-to-shelf tracking for one of retail’s most space-constrained operating environments.

Market
Hong Kong
Channels
Four key accounts and independent retail
Store formats
Supermarkets, personal care, and pharmacies
Execution focus
Facings, promotions, and new launches
96.8%Facing-recognition accuracy
94%Promotion compliance
18%Reported sales-conversion improvement
40%Faster full distribution

Customer and business context

The customer is a global consumer goods company with a multi-brand portfolio across hair care, styling, fabric care, and men’s grooming. The program covered Hong Kong’s four leading supermarket and personal-care key accounts as well as hundreds of independent pharmacies. Each channel used its own item codes, promotion cadence, and display standards, while commercial teams needed one reporting structure across them.

Hong Kong retail combined limited space with high operating speed. A product might receive only one or two facings; promotion labels changed weekly and combined multi-buy discounts, add-on offers, redemptions, and handwritten price messages; and strategic launches had only a four-week window to establish distribution and shelf presence.

The requirement was not a one-time store-recognition project. The execution system had to keep pace with weekly promotions and daily launch decisions. Results needed to map correctly to store item codes and flow into the existing commercial reporting environment so sales, category, and marketing teams could act on the same launch or campaign.

Challenge

High-density shelves left little room for recognition error

Shelf spacing in dense urban stores was extremely narrow. Products could appear in only one or two facings and were often double-stacked or compressed against similar packs. Conventional detection could merge two neighboring products into one or count a single product twice.

Weekly promotions combined price, quantity, and conditional text

Promotional labels changed frequently and could include multi-buy discounts, add-on offers, redemption terms, handwritten signs, and new labels placed over old ones. Reading a price alone was not enough; the system needed to interpret the offer and connect it to the correct item code.

New products had a four-week window to establish distribution

The first four weeks after launch determined whether a strategic product secured distribution, sufficient facings, and high-visibility placement. Monthly summaries could not show where the rollout was falling behind while there was still time to act.

Solution

Separate products within narrow, crowded facings

Implementation timeline

2025 Phase 1

Distribution and item-code mapping

The program adapted recognition for Traditional Chinese labels and dense shelves, then established complete mapping between recognized products and store item codes.

2025 Phase 2

Promotion and display standards

Multi-buy and conditional promotion recognition was connected with the commercial reporting layer and channel execution rules.

2026 Phase 3

Strategic launch tracking

Speed-to-shelf monitoring followed priority products from first-day distribution through the period when facings and shelf position stabilized.

Clobotics used high-resolution fine-grained classification to separate adjacent products in compressed displays. The workflow counted facings while reducing missed products and double counting in tightly packed shelves.

Interpret promotion text and price together

A multimodal text-price matching workflow read Traditional Chinese promotion text, price values, quantity conditions, and offer logic. It then connected the interpreted promotion with the relevant store item code instead of treating the label as an isolated OCR result.

Track new-product speed to shelf

A dedicated launch pipeline monitored daily distribution, first-week facing share, and high-visibility shelf placement by channel and area. Commercial teams could see which accounts had listed the product and where execution still required intervention.

Connect field evidence with the commercial data layer

The recognized shelf and promotion results were mapped to store and item codes and delivered to the customer’s commercial reporting environment, creating one view across execution, promotion, and product-launch measures.

Clobotics technical differentiation

Fine-grained vision for extremely dense facings

Clobotics combined high-resolution detection with fine-grained classification for similar packages placed within narrow facings. The model had to separate adjacent product boundaries rather than only identify the product family, preventing one item from being counted twice or two compressed products from being merged into one. That precision supported reliable facing share and launch-display measurement.

Interpret promotion meaning and product context together

Complex promotions could not be understood by reading one price value. The multimodal workflow parsed Traditional Chinese text, quantity conditions, unit and bundle prices, and offer logic, then used the label’s spatial relationship to map the promotion to the correct store item code. This exposed commercially important gaps such as a campaign already advertised in market while a store still displayed the previous price label.

Give new launches a dedicated speed-to-shelf pipeline

Launch tracking used a daily data flow separate from routine monthly reporting. The system followed whether a priority SKU had reached the intended store, how much initial facing share it received, and whether it occupied a high-visibility shelf position. Commercial teams could identify lagging channels during the four-week launch window rather than after it had closed.

Delivery from item-code mapping to commercial dashboards

The first phase established complete mapping between recognized products and retailer item codes, resolving the fact that the same SKU could carry different identifiers across chains. Clobotics also built a training and validation set for Traditional Chinese labels, dense facings, and channel-specific shelf conditions so later outputs entered reporting with a consistent data definition.

The second phase connected promotion recognition and display rules with the customer’s commercial dashboard. Results were organized by store, channel, SKU, and campaign instead of remaining in a separate algorithm interface. The third phase added full launch-cycle monitoring on top of the same evidence base, allowing distribution, facing, and promotion decisions to share one capture and recognition workflow.

This phased delivery increased analytical depth while keeping the frontline process stable. Stores continued to be captured through the established workflow, while headquarters received progressively richer execution measures in its familiar data environment. Models, product libraries, and promotion rules could then be updated with the weekly campaign and launch cadence without repeatedly replacing field tools.

Results

Facing-recognition accuracy reached 96.8%

The system maintained 96.8% accuracy in narrow, high-density displays where small spacing and similar packs increased the risk of merged or missed facings. This exceeded the customer’s acceptance threshold for distribution measurement.

Promotion compliance reached 94%

The program identified gaps between advertised offers and in-store labels, including stores that had not applied the current promotional price. Promotion compliance reached 94%, accompanied by a reported 18% improvement in promotional sales conversion.

Full distribution time fell from four weeks to within 2.5 weeks

Daily speed-to-shelf monitoring showed where new products had not yet reached the intended accounts, facings, or visibility. The time required to achieve full distribution decreased by approximately 40%, allowing the commercial team to act during the critical launch window.

Operating result

Dense urban retail became measurable at the pace of weekly promotions and daily launches.

The program connected shelf recognition, promotion interpretation, and new-product tracking without separating them into isolated audits. Talk to Clobotics about retail execution in high-density and promotion-led markets.