BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.2.3.1//EN TZID:Asia/Kolkata X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT UID:67@cds.iisc.ac.in DTSTART;TZID=Asia/Kolkata:20240802T113000 DTEND;TZID=Asia/Kolkata:20240802T123000 DTSTAMP:20240725T145159Z URL:https://cds.iisc.ac.in/events/mtech-research-thesis-defense-cds-02-aug ust-2024-scalable-video-data-management-and-visual-querying-for-autonomous -camera-networks/ SUMMARY:Mtech Research Thesis Defense: CDS: 02\, August 2024 "Scalable Vide o Data Management and Visual Querying for Autonomous Camera Networks" DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nMtech Research T hesis Defense\n\n\n\nSpeaker : Ms. Bharati Khanijo\nS.R. Number : 06-18-02 -10-12-19-1-17219\nTitle : "Scalable Video Data Management and Visual Quer ying for Autonomous Camera Networks"\nThesis examiner : Dr. Soma Biswas\nR esearch Supervisor: Dr. Yogesh Simmhan\nDate &\; Time : August 02\, 202 4 (Friday) at 11:30 AM\nVenue : # 202 CDS Classroom\n\n\n\nABSTRACT\n\nVid eo data has been historically known not only for its unstructured nature a nd rich semantic content but also for scalability issues in terms of stora ge and analytics. Mobile aerial platforms like drones capture such videos across space and time. Advances in computer vision and deep learning enabl e automatic extraction of rich semantic information from video data\, lead ing to applications where the stored video data can be used to study and a nalyze the world retrospectively and automatically. However\, recent resea rch has highlighted the compute-intensive nature of such Deep Neural Netwo rk (DNN) models\, e.g.\, for accurate object detection\, leading to high c omputing costs that limit their applicability for brute-force analysis of all historical videos. Also\, an efficient design of such applications oft en requires co-analysis of video data along with associated geospatial and temporal metadata\, which is a challenge.\n\nWe propose a geospatial-temp oral video query system with support for semantic queries for drone videos \, extending an existing spatial-temporal database and contemporary object detection models. We develop a heuristic to enable better reuse of semant ic object detection results obtained from different model configurations ( eg. object detection model and its input resolution) . The system further motivates the need for optimizations for retrospective semantic analysis a nd storage for drone videos\, which is addressed by our novel DDownscale m ethod and the associated ingest pipeline.\n\nPrior optimizations on semant ic querying over video data focus on static cameras from city-scale traffi c/surveillance camera networks\, often leveraging the spatial and temporal characteristics of associated videos\, which are absent in videos recorde d by mobile drone cameras. We specifically focus on two such characteristi cs of drone videos. One is that drone videos have shorter durations\, unli ke those captured by static cameras. Another is that there can be large va riations in the level of detail of information captured across a fleet of drone cameras due to differences in the resolution of the camera\, the alt itude\, and the orientation from which the videos were captured.\n\nSpecif ically\, we address the need to intelligently scale-down the spatial resol ution of videos to reduce the video storage costs and semantic query/infer encing time. However\, conventional methods of manual or profiling-based e stimation of the ideal scaling ratio are compute-intensive and/or time-con suming for such heterogeneous feeds. We propose DDownscale\, a novel metho d to dynamically select the downscale factor for a video by utilizing the information on the object size in the video. We model the downscale factor and associated drop in relative recall due to downscaling as a function o f object size in the downscaled video and demonstrated that for a given DN N model and class of interest\, DDownscale generalizes well to the evaluat ed datasets. A DDownscale inequality between the relative recall drop and the hyperparameters of the method is derived. This satisfies ~ 98% of the dynamically downscaled videos across real world video datasets\, objects o f interest and parameters. The algorithm achieves over 25% reduction in to tal object detection time and 31% reduction in storage on average compared to the baseline of storing/inferencing at the original resolution\, for d ifferent user-specified target reduction in recall values ranging from 1 - - 30%\, and 96% of the downscaled videos are within the target recall drop \, for the evaluated datasets and object detection models.\n\nA simpler sp ecification at the time of ingest of target level of detail (average groun d spatial distance) captured in the video and the harmonic mean of relativ e recall drop for the class of smallest object of interest and selected ob ject detection model was derived using the above modeling to aid in the se lection of a target level of detail. Additionally\, we develop an ingest p ipeline that reduces the time to ingest drone videos using this dynamicall y downscaling method over heterogeneous edge accelerators\, and reduce the average turnaround time to ingest data from multiple clients by ~ 66%\, d espite the downscaling time overhead\, compared to uploading original reso lution video without downscaling.\n\n\n\nALL ARE WELCOME CATEGORIES:Events,Thesis Defense END:VEVENT BEGIN:VTIMEZONE TZID:Asia/Kolkata X-LIC-LOCATION:Asia/Kolkata BEGIN:STANDARD DTSTART:20230803T113000 TZOFFSETFROM:+0530 TZOFFSETTO:+0530 TZNAME:IST END:STANDARD END:VTIMEZONE END:VCALENDAR