\r\n
\r\n
\r\n
Content Aware 360 Degree Video Optimization in Mobile Systems
\r\n
Development of an end to end system to efficiently stream 360 Degree videos. Visual content analysis including\r\n image saliency, object composition, and optical flow is incorporated to design a tiling scheme to encode the 360\r\n degree videos for efficient bandwidth utilization and improved Quality of Experience. My current contribution\r\n includes mathematical modeling of user behaviour (requested field of view) using the existing datasets.\r\n
\r\n
\r\n
\r\n
Combining object detectors for stereo images
\r\n
A geometric approach to combine object detections on corresponding images of two stereo cameras on autonomous\r\n driving vehicles, to increase precision and recall. A fasterRCNN-resnet101 detection model was used on KITTI\r\n can Cityscapes autonomous driving datasets for the experiment. LIDAR and stereo image depth calculations were\r\n done. Used Tensorflow, keras and openCV was used.\r\n
\r\n
\r\n
\r\n
Generalization ability of object detection/ segmentation models
\r\n
Observed and evaluated the performance of object detection and object segmentation models across various\r\n autonomous driving datasets of different scenes and cities. Objective is to explore the correlation between model\r\n architectures and scene properties on generalization ability. Experiments are being carried out with KITTI,\r\n Cityscapes, berkeley and Apollo scape\r\n
\r\n
\r\n
\r\n