Updates on project Meeting Behaviour & LPR
Dataset Given to me
- Train Images: 19k images
- Validation: 4K images
- Testing: 2k images
- Classes of Data
- Drinking
- Normal
- Texting
- Phone Calling
- Stretching
- Model Trained: VGG16/19
- Training Accuracy:100
- Validation:80
- Testing:63
Problems with the Given Data
- Duplicate Data: Both Training and Validation data has both duplicated data, parts of training images pasted as validation images
- Overfitting in Validation dataset: The reason above, caused overfitting in the validation data
- Less Data
Changes made to Given Data
- Duplicate data removed
- Spilt of training/validation changed
- Increased training set to 51k images
Models Trained-On the Given Dataset
| Name of the Model | Training | Validation | Testing |
|---|---|---|---|
| VGG19 | 50 | 25 | 20 |
| InceptionV3 | 90 | 45 | 25 |
| Vanilla CNN(4 Layers) | 96 | 99 | 55 |
| New Model-Vanilla CNN with more layers | 83 | 78.6 | 79.5 |
Integration of the MRCNN model
- MRCNN Model cannot be appended to the existed model due to pictures in the training set which contain both the cup and the phone
- Combined CNN+MRCNN model
- Results
LPR-Approaches taken
- OpenVino Toolkit- LPR Detection model Results
- OpenVino Toolkit- LPR text Detection
- Basic Image Segmentation+Tesseract
Basic Image Segmentation
OpenVino-Text Detection
Updates on project Meeting Behaviour
By archana iyer
