Week 6 Cecilia La Place
Progress so far... Created 5 different train-val-test splits Trained on split 1 and 2 Tested split 1 Training on split 3 Testing split 2
Papers Read Papers read Sky Region Detection in a Single Image for Autonomous Ground Robot Navigation - Shen et al Horizon Detection Using Machine Learning Techniques - Fefiltyev et al Convolution LSTM Network: A Machine Learning Approach for Precipitation Nowcasting - Shi et al
Shen et al Discusses different sky segmentation methods Need faster image processing for computers not as powerful such as small robots Highlights problems in sky segmentation Proposes method for faster image processing Preliminary information from a gradient Preliminary sky segmentation Combines the two to achieve a better sky segmentation
Fefiltyev et al Sky segmentation using machine learning classifiers SVM, J48, Naive Bayes Detects horizon lines Small dataset (20 images) Goal is to classify sky and ground pixels presuming common knowledge of where the sky is
Shi et al Nowcasting is the precise and timely prediction of weather and a problem in weather forecasting Approaches the problem from a machine learning perspective Applies fully connected LSTM which has temporal but not spatial information Proposes new model the convolution LSTM to account for the spatial information as well
Data 28 cameras for training 4 cameras for validation 13 cameras for testing On average based on mIOU split 1 had an accuracy of 43.3% The model is recognizing and trying to segment on nighttime and low light photos (better than Cityscapes!) Outperforms some cameras evaluated by Cityscapes
Split 1 - IOU and more Camera Cityscapes Refinenet 3395 49.25% 59.96% 8733 54.01% 58.80% 17218 31.70% 49.42% 1093 62.56% 39.81% 4679 58.75% 28.80%
Split 1 - Visual
Split 1 - Visual cont.
To do Gather more resources for literature review Start laying out poster for the poster presentations Train splits 4 and 5 Retrain split 2 due to overfitting Test split 2-5