White Papers

7 Retail Analytics with Malong RetailAI® on DELL EMC PowerEdge servers
products into a trace. Every trace is represented as a sequence of binary elements. The 1
st
part
in an element is a “location” within the field of view, the other part is a “timestamp” indicating
approximately when this product moved to this location.
3. The POS message handler will interface with the register messaging system to incorporate into
the processing pipeline. Each scan signal must contain when the scan happened and what
product scanned. If a customer tries to scan a wine with a wax candle barcode, then a signal
with candle barcode will be sent to the decision module.
4. The decision module will decide whether the current scan is a possible scan fraud event, e.g.
ticket-switching or mis-scan by inferring based on the following pieces of information:
i. The traces including inferred products, their movement and trajectory,
ii. The physical context of the register in the field of view,
iii. The register POS messages, including their scan timing and scan products.
Note: the decision module only considers the physical objects in the scene; that is, it in no way
whatsoever considers the people in the scene. This ensures customer privacy and no bias, a
fundamental principle in responsible AI.
All models in the pipeline are heavily optimized with NVIDIA TensorRT to best leverage the NVIDIA T4
Turing architecture Tensor Cores with mixed precision accelerated inference, which significantly
increases throughput and efficiency on the Dell EMC PowerEdge R7425. These optimizations provide for
a 480%+ speed up when compared to not using TensorRT on the same hardware. Compared to running
on CPU only, the difference with an optimized GPU version is a 99%+ reduction in processing time.
Processor
Speed of Core Model
% Difference to Optimized
CPU
4.7 seconds
-99.36%
NVIDIA T4 GPU (without optimizations)
0.175 seconds
-82.8%
NVIDIA T4 GPU with TensorRT optimizations
0.03 seconds
N/A
In terms of how many SCOs can be supported via the Malong RetailAI® solution:
1. Simultaneously, 4-6 SCOs can be supported per single T4, depending on specific
configuration settings.
2. Non-simultaneously, effectively an unlimited number. This is relevant because all the
SCOs in a store are typically not active all at once, all the time. So, in theory, all the
SCOs can be mapped to a single T4, but when the concurrency reaches maximum, a
random subset of SCOs will be covered.