Occlusion
Partial visibility can hide important product features.
Retail computer vision
When Correct Product Detection Is Not Enough
A retail shelf vision workflow that combines product detection, classification, and shelf-position processing to interpret camera observations in their operational context.
- Industry
- Retail
- AI task
- Detection + Classification
- Edge environment
- NVIDIA Jetson
- Input
- Raspberry Pi camera imagery
The problem
What the system needed to understand
The objective was to analyze retail shelf imagery, detect and classify products, and interpret how those observations related to expected shelf positions and changes in shelf state.
Retail operations need more than a list of visible objects. Each observation must be understood in relation to where the product appears and what that position represents on the shelf.
Operational context
Why detection alone wasn’t enough
Detection identifies what is visible. Shelf-state interpretation determines what that detection means relative to an expected product position.
This distinction matters when a product is partially visible, visually similar products are difficult to distinguish, or a back-row item appears through a gap in the front row.
Engineering challenges
Why shelf scenes were difficult
Depth
A back-row product can become visible through a front-row gap.
Identity
Similar packaging can make detecting a product and identifying it two separate challenges.
System architecture
From shelf image to operational context
System architecture
CaptureRaspberry Pi camera devices
TransportHTTP / API
Edge AINVIDIA Jetson
ServicePython serviceGunicorn + systemd
DetectionYOLOv8Product detection
ClassificationEfficientNet-B3Product classification
ContextShelf-position / state processing
OutputOperational output
Detection + classification
Separate models answered separate questions
Model flow
Shelf image
YOLOv8Detect product
Product region
EfficientNet-B3Classify product
Product identity
Position / state processing
Detection answered where a product was visible. Classification answered what the product was. Shelf-position processing added the operational context needed to interpret the observation.
Shelf-position processing
Position turned observations into context
Detection and classification outputs were interpreted together with expected shelf positions and changes in shelf state.
Camera visibility and shelf occupancy are not equivalent. Relating detections to position and state made that distinction explicit.
Back-row visibility
A visible product can belong to a different depth
Illustrative example · back-row visibility through a front-row gap
Before
Front rowABCD
Back rowABCD
After front product B is removed
Front rowAgapCD
Back rowAB ↑CD
Camera“I can now see B.”
Shelf state“The front position is still empty.”
Visible product≠Occupied front position
End-to-end system path
Performance extends beyond inference
Capture
Transfer
Inference
Classification
Position Processing
Application Response
End-to-end system performance depends on the complete operational path, not model inference alone.
Engineering lessons
What mattered across the pipeline
Detection confidence is not business confidence
A confident detection does not automatically determine the state of a shelf position.
Localization and identity need separate evaluation
Detection and classification failures have different causes.
Occlusion changes scene meaning
A gap can reveal a valid back-row object while the front-row position remains empty.
Position is part of inference context
Shelf observations become operationally meaningful when interpreted against expected position and state.
Edge performance is end to end
Image capture, communication, and service processing all contribute to end-to-end performance alongside model inference.
Project outcome
What the project demonstrated
The project produced an operational retail computer vision pipeline combining product detection, product classification, and shelf-position processing.
The architecture connected three distinct questions: what the camera could see, which product the models identified, and what that observation meant relative to an expected shelf position and changing shelf state.
Computer vision engineering
Facing a similar computer vision challenge?
Let’s discuss how your computer vision workflow handles the conditions it will encounter in production.