Enhancing Object Detection Performance with Data Augmentation Dataset
October 2024
Summary
Cognata’s augmented highway driving dataset allows engineers to validate perception models for ADAS and autonomous vehicles under a variety of challenging weather and lighting conditions. The dataset includes augmentations such as fog, rain, low sun, and erased lane markings, as well as a video of the original drive with class segmentation. By using the original clear-weather video as the ground truth, developers can measure key performance metrics such as precision and recall, without the need for additional real-world data collection. The benefits include faster validation cycles, cost savings, and targeted insights into which conditions challenge perception models the most. Developers can use Cognata’s DriveMatriX to generate more data tailored to these specific challenges.
Enhancing Object Detection Performance with Augmented Highway Driving Dataset
In developing perception models for ADAS (Advanced Driver-Assistance Systems) and autonomous vehicles, validating robustness under various environmental conditions is crucial. However, collecting real-world data across diverse weather and lighting scenarios can be both time-consuming and expensive, limiting the ability to thoroughly test model performance.
Cognata provides an effective solution with its augmented highway driving dataset. This dataset includes an original video recorded in clear and sunny conditions, augmented videos simulating different weather and lighting conditions like fog, rain, low sun, and erased lane markings, as well as a class segmentation video of the original drive. These augmentations allow developers to test perception models across challenging conditions without requiring additional data collection.
Dataset Overview
The dataset consists of the original video and multiple augmented versions that test perception model capabilities under different conditions:
Augmentation | Description | Use Case |
Clear Weather | Original video captured in sunny conditions. | Serves as the baseline or ground truth for validation. |
Fog | Simulates dense fog reducing visibility. | Tests model performance in low-visibility scenarios. |
Rain | Simulates rain conditions. | Assesses model robustness to environmental noise and rain interference. |
Erased Lane Marks | Removes lane markings from the road. | Validates the model’s ability to detect vehicles without clear road boundaries. |
Low Sun | Simulates glare caused by low sun. | Challenges the model’s ability to detect objects in high-glare settings. |
Class Segmentation | A video of the original recording with class-level segmentation. | Provides a visual breakdown of object classification for model validation. |
Example #1: Removed lane marking augmentation
About DriveMatriX
DriveMatriX is an advanced solution that transforms your existing test drive videos into augmented datasets by applying various weather and lighting conditions. In ADAS and AV development, real-world data collection across multiple conditions is often impractical and expensive. Most test drives occur in clear weather, limiting the variety needed for robust model validation.
DriveMatriX addresses this by generating multiple weather permutations—such as fog, rain, low sun and erased lane markings—from a single clear-weather dataset. This allows developers to validate their models across diverse conditions quickly and affordably without the need for new test drives.
The key benefits of DriveMatriX include:
- Cost Efficiency: Save time and resources by reusing existing test drives to create multiple scenarios.
- Targeted Validation: Quickly test how models perform in specific challenging conditions.
- Seamless Integration: The augmented data maintains the original annotations, allowing easy integration into existing data pipelines for validation.
- Performance Insights: The validation results highlight model weaknesses and enable targeted improvements based on specific challenges.
Example #2: Rain augmentation
Evaluate Model Performance
The primary purpose of this dataset is for validation. Developers use the original clear-weather video as the ground truth, with bounding box detections of vehicles considered 100%accurate. The performance of models tested on the augmented data can be evaluated using the following key metrics:
Precision: Measures how many of the detected vehicles are correct. Precision is calculated as:
Recall: Measures how many relevant vehicles are detected. Recall is calculated as:
A 50% confidence threshold is used, meaning any detection with a confidence score above 50% is considered valid.
Example Results from Augmented Data
The table below illustrates the performance of a YOLO object detection model validated using the augmented data generated by DriveMatriX:
Condition | Precision | Recall |
Fog | 0.97 | 0.59 |
Rain | 0.96 | 0.75 |
Low Sun | 0.94 | 0.68 |
Erased Lane Marking | 0.93 | 0.96 |
For instance, in foggy conditions, the model maintains a high precision of 0.97, meaning most detected vehicles are correct. However, the recall drops to 0.59, indicating the model struggles to detect all vehicles under low-visibility conditions. In contrast, the erased lane markings scenario shows high recall (0.96), meaning the model effectively detects vehicles even without clear road boundaries.
Example #3: Fog augmentation
Conclusion
Cognata’s augmented highway dataset provides an efficient tool for validating perception models in ADAS and AV systems under varied weather conditions. This dataset allows engineers to measure the precision and recall of their models in challenging environments, enabling more efficient validation without additional real-world data collection.
More importantly, after identifying which specific conditions—such as fog or rain—pose challenges for the model, developers can perform targeted training. Using Cognata’s DriveMatriX, additional augmented data tailored to these specific conditions can be generated, allowing for focused improvements in model performance and ensuring readiness for real-world deployment.
Example #4: Low sun augmentation