Sky Segmentation Dr Borji Aisha Urooj Cecilia La
Sky Segmentation Dr. Borji Aisha Urooj Cecilia La Place
Background Sky Segmentation in the Wild: An Empirical Analysis - Mihail et al Current vision algorithms are not effective when used on real world datasets Quality is affected by the weather, season, and time Using existing pretrained algorithms and methods proposed by Lu et al, Hoiem et al, and Tighe et al for their public and impact they calculated the accuracy via misclassification rate To improve upon the prior methods’ results, an ensemble model was developed to incorporate the results and the raw images via a recurrent CNN (r. CNN)
Goals Run the Refinenet Cityscapes model on the cameras from the Sky. Finder dataset Develop and train a model based off Refinenet to improve current sky segmentation Develop a model to predict the weather using either the newly trained Refinenet or a different model for sky segmentation followed by an LSTM
Current Results Refinenet’s Cityscapes model 45 Cameras ~ 60, 000 images m. IOU - 52. 56% Finetuning Refinenet Training it on a small dataset to make minor corrections
In Progress Steps Familiarizing ourselves with the topic (Lit Review) Download the datasets Finetune Refine. Net Clean the data Evaluate results Set up environment Run data on Refine. Net Future Steps Evaluate and compare results Propose new model End-to-end training Evaluate and compare results Write paper
- Slides: 8