Advice on AI Retail Solution

AI & Data Science Computer Vision & Image Processing
,
1

Dear NVIDIA Team,

I am PJ Pretorius, owner of OK Dana Bay, a retail grocery store in South Africa. I am reaching out to seek guidance on integrating NVIDIA’s AI technologies into our store operations.

Current Setup:

  • Store: OK Dana Bay, a small-scale retail grocery outlet.
  • Existing Equipment: Standard POS systems and surveillance cameras.
  • Upcoming Hardware: We have ordered the reComputer J3011, powered by NVIDIA’s Jetson Orin Nano, to enhance our AI capabilities.

Project Objective:

We aim to implement an AI-driven system to automate stock management, including:

  • Real-Time Stock Monitoring: Utilizing shelf cameras to track product availability.
  • Automated Notifications: Alerting staff when stock levels are low or shelves are empty.
  • Dashboard Integration: Developing an admin interface for monitoring and analytics.

Challenges:

A primary concern is efficiently training AI models to recognize our diverse product range without capturing extensive image datasets for each item.

Request for Guidance:

We seek your expertise on the following:

  1. Efficient Model Training: Recommendations on tools or methodologies to train AI models with minimal data, possibly leveraging synthetic data generation or pre-trained models.
  2. Developer Program Participation: Information on eligibility and benefits of joining the NVIDIA Developer Program, considering our status as a small retail business.
  3. Recommended Tools and SDKs: Advice on NVIDIA tools, such as Metropolis, DeepStream, or TAO Toolkit, that would best suit our project requirements.
  4. 3D Model Generation: Solutions for creating 3D models of products to facilitate synthetic data generation, especially when supplier-provided models are unavailable.

We are committed to leveraging AI to enhance our operations and believe NVIDIA’s technologies can play a pivotal role in this transformation. Any assistance or resources you can provide would be greatly appreciated.

Thank you for your time and consideration.

Best regards,

PJ Pretorius

2

Hi @pjpretorius.work

The NVIDIA Developer Program is free to join. information on what it gives you access to can be found here. Learn New Technical Skills | NVIDIA Developer

I would recommend getting onboarded and trying out some of these tools yourself to determine which ones work for you. You can also take a look at the NVIDIA Blueprints which show example workflows for a range of use cases. Try NVIDIA NIM APIs

Best,
Sophie

3

Would I be able to use 3D Models captured by a platform like Polycam to generate synthetic data to train my Retail Object Detection model?

4

I’m not familiar with Polycam, but don’t see why not - so long as their license agreement allows for such use of the data.

5

Thank you so much; I appreciate it.

My big question is, what is the most efficient way to train the Retail Object Detection model on local South African products/brands, as there is not much coverage for this yet?

Are there any methods you would suggest?

Kind Regards,

PJ Pretorius.

6

Hi @pjpretorius.work tao_tutorials/notebooks/tao_launcher_starter_kit/retail_object_detection/retail_object_detection.ipynb at main · NVIDIA/tao_tutorials · GitHub We have an example of finetuning the retail object detection model in that notebook. Thanks,

Sophie

7

I have flashed my device with Jetpack 6.2.

Which versions of CUDA, cuDNN, TensorRT and Deepstream would you suggest as I don’t want any incompatibilities?

Kind Regards,

PJ Pretorius

8

Hi,

I recently flashed my Seeed Studio ReComputer Orin J3011 with JetPack 6.2 and have been facing issues with DeepStream and missing dependencies.

System Details:

Device: Seeed ReComputer Orin Nano J3011
JetPack: 6.2 (flashed via Seeed’s official image)
CUDA: 12.6
DeepStream: 7.1
TensorRT: 10.3.0.30

Issues Faced:

  1. Missing libnvdla_compiler.so

Running deepstream-app --version gives:

error while loading shared libraries: libnvdla_compiler.so: I cannot open shared object file: No such file or directory

find / -name “libnvdla_compiler.so” returns nothing.

  1. Can’t Install nvdla-compiler orlibnvdla via APT

Running sudo apt-get install -y nvdla-compiler → E: Unable to locate package nvdla-compiler

Is this removed in JetPack 6.2? What’s the alternative installation method?

  1. Should I Reflash or Fix This Manually?

I still have the original Seeed JetPack 6.2 image. Should I reflash the device, or is there a way to fix this without a complete reset?

9

Hi @pjpretorius.work mind starting a new topic on this in the Jetson forum? That way the right people can help you out!

11

You need to do some data collection. Take photos of the SKUs and train the model with this data.. I am working on a similar use case, but from the manufacturer’s dispatch point of view. my use case is manufacturer dispatch from warehouse . We conduct object detection and counting before the items are loaded into the truck. Since the SKU brands are mainly local, you want to do get the pretrained model and training it on localized dataset.