I. Project title Sketched Object Classification (Tu-Berlin Dataset)
II. Project introduction
Convolutional neural networks are known to recognize objects by texture rather than shapes. This makes models exhibit different behavior from humans, who recognize objects based on shapes.
Sketched images hold vastly different distribution of image and hold little information about texture, which makes models trained on natural images difficult to recognize and classify into correct categories.
Training data: 14000 images Validation data: 3000 images Test data: 3000 images
III. Dataset description (need details) For training dataset, it has 20,000 sketch images across all the 250 object categories. The validation set has 1,340 images. The labels of the training and validation set is the name of the folders. The test dataset is randomly chosen of 10 images from each of the 14 categories. Therefore the test set has 140 images in total.