Enhanced Traffic Analysis Model's Accuracy by 35% through Aerial Image Annotation

The Client

A Government Agency

Our client is responsible for urban planning and development in one of the fastest-growing metropolitan areas in the US. The agency utilizes an advanced traffic analysis model that relies on aerial imagery to monitor and manage city traffic and infrastructure.

Project Requirements

Accurately Annotating Aerial Images of Traffic Based on Pre-Defined Classes

The agency sought SunTec AI's expertise to annotate over 2000 aerial images required for training their traffic analysis model. The project required precise identification and categorization of eight distinct object classes in these images, which included cars, SUVs, vans, pedestrians, motorbikes, cyclists, trucks, and buses.

Project Challenges

Addressing Image Quality and Lighting Issues for Image Annotation

While working on this project, our team encountered a few challenges, such as:

Image Quality and Consistency

The aerial images varied greatly in resolution and clarity, making it difficult to identify and distinguish diverse vehicles/objects on the roads, especially in densely populated areas.

Different Lighting Conditions

Variations in lighting across different images (from bright daylight to shaded areas) further affected the visibility of objects in the aerial shots.

Our Solution

Ensuring Precision in Labeled Datasets with a Multi-Pass Annotation Process

We employed a team of five experienced annotators who were proficient in using the client's specified image annotation tool, LabelImg. Leveraging their subject matter expertise and the bounding box annotation technique, they accurately categorized and labeled required objects in the aerial images. We adopted a multi-pronged approach to overcome the project challenges and ensure all the labeled images met the client's expected accuracy standards. Our service involved:

After

Before

Defining Annotation Criteria

We created detailed image labeling guidelines for the project, including visual references for each object class and a decision framework for handling ambiguous cases.

High Precision Zooming

We advised annotators to work at 100% zoom-in using the LabelImg image annotation tool for precise object identification and bounding box placement and then zoom out to verify context.

Feedback Sessions and Team Meetings

We held regular team meetings to discuss complex annotation cases and ensure consistency across annotators. Furthermore, our project manager maintained regular contact with the client, facilitating real-time adjustments in the labeling process by incorporating feedback.

Quality Assurance

We implemented a rigorous, multi-level QA process to ensure accuracy and consistency in the labeled dataset. From defining the list of classes for image labeling in the tool to manually verifying the annotated images, we involved our subject matter experts at every stage. The quality assurance process included:

Peer review and quality check to verify that bounding boxes were precisely placed around each object in the images, adhering to predefined classification standards.
Secondary review by senior annotators for addressing complex cases requiring more nuanced judgment, particularly for images with lower clarity or higher object density.
Final check by the project lead to ensure that the annotations conformed to all project specifications and aligned with the client's quality standards.

Project Outcomes

The Impact of Accurately Labeled Image Datasets on the Traffic Analysis Model's Performance

35% Increase in Model Accuracy

Accurately labeled image datasets enhanced the traffic analysis model's object detection accuracy, enabling better traffic monitoring.

20% Improvement in Traffic Flow Monitoring

The enhanced performance of the AI model significantly improved the agency's ability to monitor and respond to traffic flow issues, facilitating effective urban planning and congestion management initiatives.

Additionally, our annotation quality and guidelines became a benchmark for the agency's future projects, ensuring consistent quality across their data pipeline.

Improving Accuracy by 35% for an AI Traffic Analysis Model with Large-Scale Aerial Image Annotation Services