Video Annotation
Video
Object DetectionScope of Atlas Capture The Atlas Capture project involves collecting first-person perspective videos of individuals performing real-world physical tasks. These tasks include activities such as canvas assembly, object polishing, item packaging, and material measurement. The aim is to capture natural and continuous workflows instead of isolated or staged movements in order to support advanced robotics and machine learning training. Specific Data Labeling Task Performed The main responsibility involves breaking continuous video recordings into meaningful, goal-oriented segments and providing accurate descriptive annotations for each segment. The labeling process includes: 1. Segmenting videos based on changes in the task goal rather than simple hand movements. 2. Writing clear and standalone descriptions using present participle "ing" sentence structures. 3. Describing actions at the highest accurate level of abstraction while avoiding unnecessary micro-action details. 4 Identifying relevant objects and important locations without over-describing irrelevant attributes. Quality Measures Followed To maintain high quality standards for machine learning and robotics datasets, strict annotation guidelines are followed throughout the process. 1. Labels remain completely literal and only describe actions that are directly observable in the video. 2. Speculation, assumptions, or inferred intentions are strictly avoided. 3. Annotated segments are kept within the required duration range, with shorter clips merged and longer clips split at natural task boundaries. 4. Annotation language remains consistent, avoids instructional phrasing, and maintains uniform object naming across the dataset. 5. Numerical quantities are replaced with natural descriptive language where appropriate.