Data Labelling for Remotasks
Image
Bounding BoxProject Description: Autonomous Vehicle (AV) 2D/3D Scene Understanding Project Goal: To identify, classify, and annotate urban traffic elements (vehicles, pedestrians, traffic lights, lane markings) in 2D image frames and 3D LiDAR point clouds to enable AI "see" and interpret its surroundings. Tasks Performed: Bounding Box (2D): Drawing precise 2D rectangles around cars, trucks, pedestrians, and cyclists. Polygon Annotation: Outlining irregular shapes such as vehicles at angles or specialized machinery for accurate pixel-level identification. LiDAR 3D Annotation: Labeling 3D "cuboids" in point cloud data, connecting them across frames to track objects in 3D space. Attribute Annotation: Assigning specific attributes to bounding boxes (e.g., vehicle type, lighting status—on/off, pedestrian action—walking/standing). Semantic Segmentation (Masking): Painting pixels of the road, sidewalk, and lane lines to distinguish drivable surfaces. Project Size Volume: These projects often handle thousands of images per batch. Scale: Projects are continuous, with large teams of "Remotaskers" working concurrently to ensure high-volume output (e.g., millions of annotated objects). Duration: Ranging from temporary, high-priority batches to long-term, ongoing efforts. Quality Measures Adhered To Remotasks maintains high accuracy using a multi-step quality assurance system: Reviewer System: Submitted work is checked by senior reviewers or "reviewers". Accuracy Thresholds: Annotators must maintain a high accuracy percentage (e.g., often >70-80% required) to remain enabled on a project. Consensus/Consolidation: Similar images are often sent to multiple users; the platform validates the final answer based on the consensus, discarding outliers. Strict Guidelines (Pedantry): Adherence to detailed, project-specific PDF instruction manuals regarding pixel-tight boxes and strict labeling rules. Ground Truth Tracking: Regular audit tasks are inserted to ensure annotators are producing accurate, reliable data