The pre-trained CV AI Object Detection model was tested with images of "Pallet Jacks" from the Logistic Objects in Context (LOCO) dataset to visualize real-world model performance. We tested inference on real-world images from the LOCO dataset with the model accurately detecting "Pallet Jacks" objects. The following figure shows "Pallet Jacks" objects detected (delineated with red bounding boxes) in real-world images.
Findings
- Average CPU and GPU utilization of approximately 25% and approximately 72% respectively during model training.
- The CV AI model was solely trained on synthetic data. For further model optimization, it would be interesting to conduct training with both synthetic and user-curated real images.
Note on resource assignment
The above test cases (3D Simulations, Synthetic Data Generation, and Computer Vision AI Model Training) were run sequentially. In practice during Digital Twin development, these workflows might be required to run concurrently. We found that a virtualized NVIDIA Omniverse Enterprise deployment allowed for flexible resource assignments.
The following figure shows a snapshot of L40 utilization for concurrent workflows (3D model development, SDG generation, and CV AI model training) running on a single Dell PowerEdge 760xa.
