Deep Learning for Computer Vision

Task 1: Image Classification with CNNs

Dataset and Objective: We created a model to predict the breed of a dog or cat, given an input image. The dataset can be downloaded from here.
Models: We started off with a self-made CNN-based architecture, using nn.Conv2d, nn.MaxPool2d, and nn.ReLU. After training and evaluating the model, we looked into adapting and finetuning pretrained models like InceptionNet, ResNet, EfficientNet, and MobileNet.
Evaluation: We evaluated each model on the basis of accuracy, precision, recall and F1-scores.

Dataset and Objective: We created a model that outputs a segmentation mask of people inside an image. this can be used for blurring a person in an image, for example. We used the dataset found here.
Models: We started off with a simple convolution-based autoencoder and created a sequence of nn.Conv2d and nn.MaxPool2d layers that output a latent vector. We fed this into a sequence of nn.ConvTranspose2d and pooling layers that output a segmentation mask at the end.
Evaluation: We plotted a few examples of our output segmentation map and compared it to the ground truth. we used the IoU and DICE coefficient to evalute our model.

Dataset and Objective: We created an LSTM model to caption images. The dataset can be found here.
Models: The code explains the model used.
Evaluation: We loaded our own examples to test the model.

Dataset and Objective: We used the same dataset as in Task 1 with the same objective of classifying pet breeds based on an input image.
Models: We used a vision transformer.
Evaluation: We compared the number of parameters and the same metrics with the models from Task 1.