page-flip-detector
Data Description:
We collected page-flipping videos from smartphones and labeled them as flipping and not flipping.
We clipped the videos as short videos and labeled them as flipping or not flipping. The extracted frames are then saved to disk in sequential order with the following naming structure: VideoID_FrameNumber
Goal(s):
Predict if the page is being flipped using a single image.
Success Metrics:
Evaluate model performance based on F1 score, the higher the better.
Project Development
Highlights
- Created a custom earlystopping class to save the best checkpoints weitghs states.
- Created custom model with nn class from Pytorch
- Used transfer learning technique with the light model of mobile Net v2
- Achieved best results with the custom model over the mobile net
Image Classification with PyTorch
This Jupyter Notebook contains the code for training and evaluating image classification models using PyTorch.
Development
The notebook starts by importing the necessary libraries and loading the dataset. The dataset consists of images of pages being flipped or not, which are split into training and validation sets. The notebook then defines and trains two different models: cnn_model and mobilenet_v2. The first is built from scratch using the Pytorch nn module. The Mobile Net is a well-known mobile and light model, and we apply transfer learning to it. After training both models on our dataset, we found that cnn_model performed better than mobilenet_v2, achieving an F1 score of 97.5%. This indicates that cnn_model is a good candidate for further testing and deployment.
Conclusion
In this phase of testing, we trained and evaluated three different models: cnn_model, MobileNet, and ResNet. After training all three models on our dataset, we found that cnn_model performed the best, achieving an F1 score of 97.5%. However, MobileNet and ResNet also performed well, achieving F1 scores of 96.6% and 91.8%, respectively.
These results indicate that all three models are good candidates for further testing and deployment. However, the task wanted the model to be applied on mobile applications, which often means that the model has to be smaller then 40 MB.
Our custom CNN model got Estimated Total Size (MB): 51.09;
Although ResNet18 is a popular and well performing model, it is not the best choice for mobile applications. ResNet18 got Estimated Total Size (MB): 81.11;
MobileNetV2 is our choice: it is a small and efficient model that is well suited for mobile applications. MobileNetV2 got Estimated Total Size (MB): 24.88.
For future work, I intend to tweak on the custom model precision point using quantization techniques to reduze its size and try to fit in mobile applications.