SUTD 50.035 Computer Vision Coursework using Tensorflow. For more information, refer to https://istd.sutd.edu.sg/undergraduate/courses/50035-computer-vision.
Part 1 (OpenCV exercises) - load images, transformations, changing color space, 2D convolution (image filtering), average filter, gaussian blurring, median filter, image gradient, histogram equilization, threshold (simple & adaptive)
Part 2 (k-nearest neighbor & softmax classifier) - train a knn classifier and softmax classifier in Tensorflow
2. HW2 [Code] - Part 1 (Tensorflow exercises), Part 2 (Manual implementation of CNN layers in Tensorflow)
Part 1 (Tensorflow exercises) - create CIFAR10 dataset, initialise and train two/three-layer networks
Part 2 (Manual forward/backward implementation of CNN layers) - affine, ReLU, softmax, convolution, max-pooling, dropout
3. HW3 [Code] - Generative Adversarial Networks (GAN), Style Transfer, Feature Inversion, Network Visualisation in Tensorflow
(i) Generative Adversarial Networks - Vanilla GAN, Least Squares GAN (LS-GAN), Deep Convolutional GAN (DCGAN) on MNIST dataset
| Vanilla GAN | Least Squares GAN | Deep Convolutionl GAN |
|---|---|---|
 |
 |
 |
(ii) Style Transfer - Implementation of "Image Style Transfer using Convolutional Neural Networks" (Gatys et al., CVPR 2015)
| original input image | style source image | style transferred image |
|---|---|---|
 |
 |
 |
|
 |
 |
Feature Inversion - Implementation of "Understanding Deep Image Representations by Inverting them" (Aravindh Mahendran et al., CVPR 2015)
| original input image | style source image | feature inversion image |
|---|---|---|
|
|
 |
(iii) Network Visualisation - Using image gradients to generate new images
(a) Saliency Maps: Saliency maps are a quick way to tell which part of the image influenced the classification decision made by the network.
| saliency map |
|---|
 |
(b) Fooling Images: We can perturb an input image so that it appears the same to humans, but will be misclassified by the pretrained network.
| fooling images |
|---|
 |
(c) Class Visualization: We can synthesize an image to maximize the classification score of a particular class; this can give us some sense of what the network is looking for when it classifies images of that class.
| recreational vehicle | tarantula |
|---|---|
 |
 |