This document compares the inference times of two different implementations for a series of images. The first set of data represents inference times using Ultralytics code, and the second set represents inference times using our own implementation.
| Image Number | Path | Ultralytics Inference Time (ms) | Our Code Inference Time (ms) |
|---|---|---|---|
| 1 | C:\Users\thata\intern\code\pre-built-models\assets\000000000009.jpg | 90.2 | 92.37 |
| 2 | C:\Users\thata\intern\code\pre-built-models\assets\000000000025.jpg | 80.1 | 89.04 |
| 3 | C:\Users\thata\intern\code\pre-built-models\assets\000000000030.jpg | 10.0 | 5.00 |
| 4 | C:\Users\thata\intern\code\pre-built-models\assets\000000000034.jpg | 10.0 | 5.27 |
| 5 | C:\Users\thata\intern\code\pre-built-models\assets\000000000036.jpg | 90.1 | 80.48 |
| 6 | C:\Users\thata\intern\code\pre-built-models\assets\000000000042.jpg | 10.0 | 6.00 |
| 7 | C:\Users\thata\intern\code\pre-built-models\assets\000000000049.jpg | 10.0 | 6.95 |
| 8 | C:\Users\thata\intern\code\pre-built-models\assets\000000000061.jpg | 80.2 | 82.79 |
| 9 | C:\Users\thata\intern\code\pre-built-models\assets\bus.jpg | 80.0 | 86.33 |
| 10 | C:\Users\thata\intern\code\pre-built-models\assets\zidane.jpg | 80.2 | 89.60 |
- Ultralytics Code Total Inference Time: 3.32 seconds
- Our Code Total Inference Time: 2.09 seconds
- Preprocessing Time: 1.0 ms per image
- Inference Time: 54.1 ms per image
- Postprocessing Time: 4.0 ms per image
Total Average Processing Time per Image: 59.1 ms
-
Individual Inference Times:
- Inference times for individual images vary between the two implementations. Our code generally shows lower inference times for some images compared to Ultralytics.
-
Total Inference Time:
- Our code demonstrates a faster total inference time (2.09 seconds) compared to the Ultralytics code (3.32 seconds).
-
Complexity and Overhead:
- The Ultralytics code includes a complex saving method that expects a specific module structure, contributing to additional overhead time.
- Our code avoids this complexity by not requiring the entire module structure, resulting in reduced overhead and faster inference times.