Project Overview:

My partner and I developed a model and data loader to detect pet noses as key points in images from the Oxford-IIIT pet dataset. We customized the ResNet18 model from PyTorch by adding an average pooling layer and then a fully connected layer to tailor the model to our needs, leveraging ResNet18's reliability for optimal key point detection using pre-trained weights. The model demonstrated solid accuracy, with a mean absolute error of 26.11, indicating that the predictions differed from the actual values by 26.11 units. Qualitative results further confirmed the model's success in pinpointing pet nose locations. I was responsible for programming the data loader in Python, retrieving ground truth values from a plain text file, preparing the tensors for training, and writing the code to evaluate and test the model.

Project Reflection:

I thoroughly enjoyed working on this machine vision with deep learning project. It was gratifying to visualize the qualitative results of the model. If I were to continue this project, I would experiment with other deep learning models suited for key point detection tasks. The goal would be to lower the mean absolute error by experimenting with various models and tuning hyperparameters such as epochs, batch size, optimizer, learning rate, device, loss function, and scheduler to find the optimal combination for enhancing the model's accuracy and reliability.

I look forward to working on more machine vision tasks, exploring the capabilities of OpenCV, and investigating other practical applications of machine vision.

Key Skills:

Python, Machine Vision, Pytorch, Deep learning, AI model development, Keypoint Detection tasks, Debugging, Tensors, Data Extraction