14,272 traffic light boxes in 3 days: how LabelFort annotated 4,564 images for perception models

The client builds traffic light detection and autonomous vehicle perception systems that depend on precise object localization - especially for small objects that occupy limited pixels and appear under varying lighting and weather conditions. Their models require tight, one-object-per-box annotations with standardized coordinates, not loose regions that include poles, wires, or adjacent objects.

LabelFort processed 4,564 traffic light images and delivered 14,272 labeled bounding boxes within 3 days, using a single-pass workflow across 29 annotators with strict tight-box guidelines on every instance.

The challenge

Traffic light bounding box annotation requires high precision. Traffic lights are often small, frequently overlap with poles and wires, and may be partially truncated at image borders. Images are captured under varying lighting and weather that reduce visibility.

Even minor errors - loose boxes or multiple objects in one rectangle - reduce model performance. The client needed consistent tight boxing and standardized coordinate formats at speed.

• Small objects occupying limited pixels in each frame
• Overlap with poles, wires, and surrounding infrastructure
• Partial truncation at image borders requiring visible-portion-only boxing
• Varying lighting and weather reducing visibility
• One traffic light per box - never two objects in a single rectangle

What we delivered

Each image was annotated with bounding boxes for every visible traffic light. Images could contain multiple traffic lights, and each instance received its own separate box.

Bounding box rules enforced:

• Tight box rule - boxes tightly touch object edges with minimal background
• One object = one box - never include two objects in a single rectangle
• Object fully inside - all visible parts must be inside the rectangle
• Truncated objects - annotate only the visible portion when cut by the image border
• Overlapping objects - boxes may overlap, but each traffic light gets its own box

Our approach

LabelFort ran the engagement as a single-pass workflow with checklist-driven quality rules applied across the full annotator team.

• Annotators drew bounding boxes around every traffic light following tight-fit rules.
• Truncated objects were boxed on visible portions only.
• Overlapping traffic lights each received independent boxes.
• Standardized coordinate formatting applied across all 4,564 images.

Results

The engagement delivered clean, model-ready object detection annotations on schedule - tight boxing standards applied across the full dataset.

• 4,564 traffic light images annotated
• 14,272 traffic lights labeled with bounding boxes
• 143 hours 10 minutes of total agent effort
• 29 annotators on a single-pass workflow
• 14,272 total labeling decisions with standardized coordinates

Key insights

• Tight box precision is critical for small-object detection like traffic lights.
• Correct handling of truncated and overlapping objects improves model generalization.
• Consistent coordinate formatting prevents training errors in object detection pipelines.
• Checklist-driven quality rules reduce noise and improve dataset reliability.

Impact

• Traffic light detection and recognition models
• Autonomous vehicle perception systems
• Smart city surveillance and traffic monitoring solutions
• Computer vision training for small-object localization tasks

Key takeaways

• Delivered 14,272 tight bounding boxes across 4,564 images in 3 days.
• One-object-per-box rules kept training signal clean for small-object detection.
• Truncation and overlap handling improved generalization across varied capture conditions.
• Standardized coordinate formatting held 29 annotators to consistent quality at volume.