Embedded Files
Project Website Link: https://sites.google.com/andrew.cmu.edu/group18website/
Members: Ruimeng Chang (ruimengc), Nidhi Jain (nidhij1), Zixiao Xing (zixiaox), Rohan Zeng (rzeng)
Problem Overview
Fig. 1 In this project, we are trying to derain images. Our idea is to use a two-stage model to detect the raindrop masks and inpaint the masked-out images to generate images without rain.
Background + Motivation
Cameras are a common method of getting visual data for many systems. They are used in many outdoor settings, such as with security cameras, autonomous vehicles, drones, and more. Oftentimes when training visual models on camera images, the training data consist of clear images of the scenery, objects, or other interesting features. However, adverse weather conditions cause the degradation of collected image quality for outdoor vision systems, thus affecting the performance of subsequent vision tasks such as object detection and segmentation. The purpose of rain removal is to restore the clean images and ensure vision system performance.
Fig. 2 On the left are two images showing the results of Google Vision's API when performing object detection in an image. The top is the original image with rain, the bottom is the image with rain removed. It is clear that removing the rain improves the performance of the API, as more object classes related to the main object in the scene are identified with greater confidence than in the original image with rain. [1]
Page updated
Report abuse