Hi everyone,
I’ve recently published a synthetic dataset focused on a specific autonomous driving edge case: Pedestrians in heavy night fog. As we know, real-world data for extreme weather is often hard to collect and label. To address this, I used Unity Perception to simulate a night-time environment from the perspective of a vehicle (specifically a Kia Ray, with a camera height of 1.4m).
Key Features of this Dataset:
-
Environment: High-density fog with low-light night conditions.
-
Quantity: 250 pairs of images and YOLOv8 formatted labels (Sample version).
-
Perspective: Realistic dashcam POV (Point of View).
-
Variety: Randomized pedestrian placements and fog densities.
I would highly appreciate your feedback on:
-
Visual Fidelity: Does the synthetic fog and lighting look realistic enough to bridge the “Sim-to-Real” gap?
-
Label Accuracy: Are the bounding box annotations precise enough for your training pipelines?
-
Data Value: How useful do you find these types of synthetic edge cases for improving model robustness?
Dataset Link: JTSGRIT/NightFogPedestrianDataset_KiaRayPOV · Datasets at Hugging Face
I’m planning to expand this dataset to include more weather conditions (snow, heavy rain) and more classes (cyclists, motorcycles). Looking forward to your professional insights!
Best regards,