vlambda博客
学习文章列表

IDEA: Data Augmentation

• Description: 
  Deep learning networks are hungry for data and data augmentation is one of the easiest ways to increase data variation. Augmentation could be as simple image flipping, cropping and scaling on up to more complicated transformations such style transfer using another deep learning network. For computer vision problems, OpenCV is often used for reading images in most of training scenarios, so why we’d like to enhance data reading with simple to use data augmentation techniques as well.

• Expected Outcomes:

• These should in particular show use with PyTorch and TensorFlow.

• Provide an API to apply single transformations to an Image or batch of Images, Rectangles (i.e. for ground truth for object detection), Masks.

• Let users combine different transformations in the class object which can apply them with some probability.

• Custom data transformations which can be included in the augmentation classes.

• Things that help with data augmentation for training networks

• Lighting functions

• spherical or cylindrical views around a planar object

• noise …

• for 3D point clouds

1. Analyze which image transformations are widely used for image classification, object detection, semantic and instance segmentation problems.

2. Create a new OpenCV’s module (or use an existing one such datasets or dnn?) with at least the following functionality:

3. Write tutorials targeting on Python wrappers due it’s the most popular language supported by different DL frameworks right now.

• Skills Required: 
  Experience in image processing and deep learning networks training for computer vision problems.

• Possible Mentors:  Intel OpenCV team member

• Difficulty: Medium to Hard

  
  

感兴趣的同学可发邮件至[email protected]与我们沟通项目细节和方案，标题请注明暑期2020+姓名+申请项目。具体报名流程请参考。