Precise Video Annotation Boosts Drone Object Detection Accuracy by 30%

The Client

A Leading Technology Company Specializing in Drone-Based Aerial Surveillance and Data Collection

Based in the US, our client offers drone-based security and surveillance support to businesses across diverse industries, including agriculture, construction, and real estate. Their solutions include high-resolution imaging, real-time data collection, classification, & analysis, and comprehensive reporting.

Project Requirements

Accurate Labeling of Drone Footage for Precise Object Detection

The client required video labeling support to prepare a training dataset for their object detection algorithm for classifying drones in flight. The goal was to enable the algorithm to identify drones at different altitudes, in varying lighting conditions, and during all possible flight stages. The footage, captured from other drones using standard and infrared cameras at different times of the day and night, included images shot in low-light conditions and during high-speed drone movement. Precise annotation of each frame was crucial to ensure the algorithm could accurately recognize drones under diverse operational scenarios.

Project Challenges

Thermal Signatures and Unpredictable Movements

The client had a large volume of drone footage (100,000 frames, equal to 55 hours of video) that had to be annotated. However, several challenges made labeling these videos complex, such as:

Thermal Signatures

Annotating drone footage from infrared cameras was difficult due to varying thermal signatures and opacity issues.

Low-Light Conditions

Videos shot during night or in low-light conditions required enhanced visibility for accurate annotation.

Drone Positioning

The drones moved unpredictably in multiple directions—up, down, forward, and backward—making it difficult for automated tracking algorithms to maintain a consistent lock on the drone's position.

Distance from Camera

Drones often appeared too far from the camera, making it difficult to capture clear aerial shots and complicating the labeling process. Drone videos required manual frame adjustments before they could be annotated.

Our Solution

Accurate Aerial Video Labeling through Manual Frame Adjustment and Bounding Box Annotation

To overcome these challenges, we deployed a dedicated team of 20 data annotators, specializing in labeling aerial footage using the client-specified annotation tool - CVAT. Our approach involved:

After

Before

Enhancing Frame Opacity

For infrared footage, our team adjusted the opacity of frames to improve visibility for accurate image labeling.

Manual Frame Adjustments

To compensate for unpredictable drone movements, our annotators carefully analyzed each frame and made manual adjustments by drawing bounding boxes around the drones. Our team also assigned unique identifiers for accurate tracking of individual drones consistently across frames.

Collaborative Feedback Loop

We maintained an iterative feedback cycle with the client, allowing for real-time adjustments and refinements based on their evolving needs and insights.

Quality Control

We established a multi-level QA process to ensure annotation accuracy and consistency. The quality control process involved:

Initial Review: Each annotated frame was initially reviewed by senior annotators to verify the accuracy of bounding boxes and identifiers.
Secondary Review: A secondary QA team conducted random checks on annotated frames to identify and rectify any inconsistencies.
Final Approval: Project managers conducted a final review to ensure that all annotations met the client's specifications and quality standards.

Maximizing the Quality of Training Dataset with Human Oversight in Annotation

Our human-in-the-loop approach was instrumental in this project. It combined the efficiency of the video annotation tool with the precision of human oversight. By involving subject matter experts in the process, we created reliable training datasets for the client's object detection algorithm.

Project Outcomes

Improved Algorithm Performance

The accuracy of the client's object detection algorithms increased by 30%, enhancing the precision of the drone surveillance system.

Enhanced Operational Efficiency

With high-quality training datasets, the client witnessed a 20% increase in overall operational efficiency.

Expanded Drone Tracking Capabilities

The client was able to extend their drone detection and tracking capabilities to challenging low-light scenarios leveraging precisely labeled infrared images.

Maximizing Efficiency in Drone Surveillance through Accurate Video Annotation