The first class in the Computer Vision training starts off by giving a good introduction on what Computer Vision is, its applications, and its relation to the human visual system. We then move to the theoretical part in which we learn how digital images are represented and processed using basic operations such as thresholding, convolutions, and filters.

These concepts will come together in the second part of the training, in which we discuss Canny Edge Detection and its relation to concepts such as Gaussian kernels, gradient computation, non-max-suppression, and hysteresis. We then dive into Scale Invariant Feature Transformation (SIFT), a powerful technique for identifying key feature points, often used in image stitching and conclude by providing a list of commonly used Computer Vision tools and resources.

The training includes theory, demos, and hands-on exercises.

After this training you will have gained knowledge about:

• Computer Vision and its applications
• Reading digital images and their representation
• Transformation operations such as thresholding
• Filters and convolutions
• Gradients computation
• Non-max-suppression
• Canny Edge Detection
• Scale Invariant Feature Transform (SIFT)
• Computer Vision tools

We start with introducing the computer vision tasks (image classification, object detection, …) and motivation for the convolutional neural network. Then we discuss the building blocks of and best practices necessary to successfully train the convolutional nets, talk about more advanced tricks to improve the accuracy. After that we introduce the classical neural net architectures that won the ImageNet image classification competitions. Finally we talk about the most recent convolutional net architectures, for object detection and segmentation tasks.

After the theory, we do a two-part lab exercise. First implement the convolutional layer from scratch in Python, then build and train a convolutional network in TensotFlow using the CIFAR-10 dataset.

The training includes theory and hands-on exercises.

After this training you will have gained knowledge about:

• Building blocks of the convolutional neural network
• Convolutional layer
• Activation function
• Pooling
• Softmax layer
• Batch normalisation and its alternatives
• Best practices and advanced tricks for training convolutional neural networks
• Transfer learning
• Normalization
• Weight initialisation
• Regularization
• Data augmentation
• Ensemble of models
• Classical neural net architectures for image classification (AlexNet, VGG, GoogLeNet, ResNet)
• Most recent neural net architectures for object detection and instance segmentation (R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNN, YOLO, SDD, RetinaNet, Feature Pyramid Network)
• Building and training a convolutional net in practice, using TensorFlow