Embedded Files
Applications of Image Gradients
Applications of Image Gradients
- Identified spatial support of particular classes in input images using saliency maps.
- Generated adversarial examples to illustrate blind spots in the network manifold that hinder generalization. In the image below, we start out with an image class of hay and perform gradient ascent to perturb the original image such that the network misclassifies it as a stingray.
- Numerically generated images from random noise using gradient ascent on target classes to visualize class models learnt by SqueezeNet trained on ImageNet.
- Implemented Neural Style Transfer by Gatys et al. using PyTorch
Content Source Image
Style Source Image
Result
Content Source Image
Style Source Image
Result
Content Source Image
Style Source Image
Result
Face Detection With a Sliding Window
Face Detection With a Sliding Window
- Computed a SIFT-like Histogram of Gradients feature representation to train a Support Vector Machine for face detection
- Boosted accuracy using hard negative mining and non-maximum suppression
- Achieved an average precision of 0.899 on the CMU + MIT test set
- The report can be found here
Mesh Simplification
Mesh Simplification
- Implemented iterative vertex pair contractions to minimize quadric error metrics as described in Garland and Heckbert for producing high quality approximations of polygonal meshes using the DGP toolkit.
- Here is a video of the results: Stanford Bunny, Torus
- Homer Simpson mesh file taken from this link.
Page updated
Google Sites
Report abuse