Computer Vision And Deep Learning

• Lab1
  This is an introductory session to Python and a few libraries that are frequently used in this course (numpy, matplotlib, opencv, keras).
• Lab2
  Implementation from scratch of a softmax classifier, trained on CIFAR-10.
  

  Weights visualization
  

  
  Accuracy: 41.65% on validation set
• Lab3
  Convolution implementation and practice with tensorflow on CIFAR-10.
  • serialization
  • use ReLu activations and He initializer
  • use regularization
  • use dropout
  • use cutout (custom layer implementation)
  Accuracy: 79.3% on validation set
• Lab4
  "The main objective of this laboratory is to familiarize you with the training process of a neural network. More specifically, you'll follow this "recipe" for training neural networks proposed by Andrew Karpathy. You'll go through all the steps of training, data preparation, debugging, hyper-parameter tuning.

  In the second part of the laboratory, you'll experiment with transfer learning and fine-tuning. Transfer learning is a concept from machine learning which allows you to reuse the knowledge gained while solving a problem (in our case the CNN weights) and applying it to solve a similar problem. This is useful when you are facing a classification problem with a small training dataset."
  

  • custom data generator
  • dataset used: GTSRB - German Traffic Sign Recognition Benchmark
  • experiment with ResNet blocks
  • transfer learning and fine-tuning with MobileNet as base model
  • data augmentation
  • custom implementation of cosine annealing scheduler
  • top 3 ensemble: 98.33% accuracy on validation set

• Lab5
  "In this laboratory we'll work with a semantic segmentation model. The task of semantic segmentation implies the labeling/classification of all the pixels in the input image.

  You'll build and train a fully convolutional neural network inspired by U-Net. Also, you will learn about how you can use various callbacks during the training of your model.

  Finally, you'll implement several metrics suitable for evaluating segmentation models."

  • image segmentation on OxfordPets dataset
  • U-Net downsample and upsample path
  • skip connection
  • checkpoints, terminate on NaN, early stopping
  • mean pixel accuracy: 92.32%
  • intersection over union: 70.28%
  • frequency weighted intersection over union: 86.83%

  

• Lab6
  "Visualizing what neural networks learn"
  

  • Display the filters in the first and second convolutional layer
    

  • Saliency map via image occlusions
    

  • Saliency maps via image derivative
    

  • Deep-dream or how to make a network hallucinate
    

Babeș-Bolyai University
Faculty of Mathematics and Computer Science
Computer Vision and Deep Learning course
Third year