Image annotator
Image
Object DetectionProject Scope The project involved annotating a large dataset of real-world street-level and indoor images to train an object detection model for a retail automation client. The goal was to enable the model to recognize product types, shelf positions, and customer interactions in store environments. Annotations needed to be highly accurate to support use cases like inventory tracking and checkout-free shopping. Specific Data Labeling Tasks Performed · Bounding Box Annotation: Drawing tight, axis-aligned boxes around individual product units (e.g., soda cans, snack bags, personal care items) · Attribute Labeling: Assigning attributes to each box, including product category, brand, size (small/medium/large), and occlusion level (none/partial/heavy) · Instance Segmentation (Polygons): Tracing exact boundaries for overlapping or irregularly shaped products where bounding boxes would include background pixels · Relationship Tagging: Marking whether multiple instances belonged to the same product family or were arranged in a visible row · Review & Correction: Spot-checking peer annotations and flagging guideline violations for rework Project Size · Total images annotated: ~15,000 (split across 45 annotators) · Annotations per image: Average of 12–25 bounding boxes + attributes · Total annotations produced: Approximately 250,000 labeled objects · Project duration: 6 weeks (phased delivery with weekly milestones) · My personal contribution: ~500 images, ~8,500 annotations Quality Measures Adhered To · Strict guideline adherence: Followed a 25-page annotation rulebook with visual examples of edge cases (e.g., partially visible products, reflections on packaging) · Blind validation set: 5% of each batch was pre-annotated by a senior reviewer; annotators whose validation accuracy fell below 95% received retraining · Two-pass review system: Each annotation passed through automated format checks, then manual spot review by a team lead · Weekly calibration tests: All annotators re-annotated a small shared image set to measure and align inter-annotator agreement (target IAA: ≥96%) · Revision loops: Flagged errors were returned within 24 hours with explanations, requiring correction before new tasks were assigned · Self-QC checklist: Required a pre-submit review against common error types (missed objects, loose bounding boxes, wrong attributes)