In this project, I leveraged the backbone of HRNet Architecture to implement a DL model for human placeta tissue (hyperspectral image dataset) segmentation. I trained the model with two options: 38 wavelength bands of spectral images and 3 bandes of spectral images after applying PCA. I evaluated the model with various metrics: mIoU, Accuracy, Precision and Recall. The mIoU score of the trained model is 89% on the test dataset.
In this project, me, David and Milan, who are my teammates, performed a domain adaptation module to dehaze nighttime hazy images. Milan developed a state-of-the-art GASDA algorithm to train on real and hazy images in order to obtain the depth of images, I implemented the CycleGAN to train domain adaption on our dataset (synthetic and real images) and David managed to write scripts for collecting real nighttime hazy images and implement different metrics for model evaluation such as MSE, PSNR, SSIM, CIEDE2000, BRISQUE, FADE. Our result outperformed existing state-of-the-art approaches in term of real nighttime image dehazing but the result for nighttime synthetic image dehazing was not good and needed for further research.
In this project, I presented and analyzed the mathematical background of two different clustering algorithms: K-Means, Agglomerative algorithms. I implemented these approaches on Google Colab to conduct experiments with the MoCap Hand Postures Dataset from the Center for Machine Learning and Intelligent Systems from the University of California, Irvine. For model evaluation, I developed and compared various metrics such as Homogeneity, Completeness, V_measure, Fowlkes_Mallows, Silhouette, Hungarian Error, Entropy, Gini Coefficient . Addtionally, I also implemented SVM, K-NN, Decision Tree Classifier and CNN-based approach to compare the difference between clustering and classificaiton problems.
In this project, me and Milan, who is my teammate, modified OpenPose algorithms to work with the dataset of Pig. We replaced the backbone by the EfficientNet architecture along with spatial attention mechanism to extract features. The result was evaluated on AP_OKS and AR_OKS and the best feature-extractor backbone was selected. This project is under writing for academic conference submission.
In this project, I reviewed 22 papers on the topic top-down approahces on human pose estimation. I deeply discussed about the mathematical background of each method and summarized shortly in order to provide newcomers an extensive review of deep learning methodsbased 2D images for recognizing the pose of people. The papers, taken into account, were between 2016 and 2020.