DETR Model Shows Different Results on Different Hardware after Optimized with OpenVINO™

1. Optimized DETR model with model optimization. 
2. Ran benchmark_app with the optimized model (DETR model) on the SPR machine and saved the latency and throughput.
3. Ran the same benchmark_app and model on the Icelake machine and saved the latency and throughput.
4. Both results showed significant differences.

Below is the resolution after the details are examined.

1. The accuracy difference in different hardware is expected behavior due to the memory and core of the processor.
2. This can be supported by documentation where some comparisons were made on different hardware.
3. For object detection, it is also expected to have multiple detections with different probabilities.
4. In the object detection sample, the developer added a filter called confident_threshold to filter out detection with a lower probability.
5. Suggestion to perform filter out the lower score in the code.