- Traffic Images downloaded from API and annotated using Roboflow, which is then exported for training of model
- Import pre-trained model forom YoloV8 and fined-tuned (train) with exported data, then exported in ONNX Format
- Prepare ONNX model for inference
Tools | Description |
---|---|
Roboflow | Computer vision framework for data collection, annotation, preprocessing and augementation. Supports custom datasets and smart labeling with pre-trained model on COCO (Common Objects in Context). |
YoloV8 | Pre-trained model that comes with different weights and sparcity of model. Model can be imported and fined-tuned with custom dataset for specific use cases. |
ONNX | Intermediary Machine Learning framework to convert between different ML frameworks to deploy accross differnet platforms. ONNX was used for easy deployment to Render cloud hosting & for comparison of model in different frameworks for a common platform. |
Endpoint enacted that accepts a image URL. An example of the annoted image is shown below:
Confidence values for identified objects are also logged:
Class 0 : Score 0.906679630279541
Class 0 : Score 0.8510133028030396
Class 0 : Score 0.8255020380020142
Class 0 : Score 0.7501163482666016
Class 0 : Score 0.7322111129760742
Class 0 : Score 0.7169588804244995
Class 0 : Score 0.7063149213790894
Class 0 : Score 0.6881899833679199
Class 0 : Score 0.6813197135925293
Class 0 : Score 0.6607131361961365
Class 0 : Score 0.6455743312835693
Class 0 : Score 0.6336331367492676
Class 0 : Score 0.3735249638557434
Testing the ONNX model on test data yields an accuracy score of 0.706.
Images are hand annotated to ensure correct classfication of objects and control over classes.