מושגי ליבה
Pit30M is a large-scale dataset that provides accurate ground truth for global localization of self-driving vehicles, enabling systematic evaluation of retrieval-based localization approaches at city scale.
תקציר
The authors introduce Pit30M, a novel large-scale dataset for image and LiDAR-based global localization in the context of self-driving vehicles. The dataset covers over 25,000 km and 1,500 hours of driving in the Pittsburgh metropolitan area, with over 30 million accurately localized sensor readings (within 10 cm error).
The key highlights of the dataset are:
Diversity: Pit30M captures diverse conditions including seasons, weather, time of day, traffic, and construction.
Scale: The dataset covers an entire city, spanning an area of around 50 km^2.
Accurate Ground Truth: The localization ground truth is obtained through a commercial batch optimization system, providing sub-meter accuracy.
Metadata: The dataset is annotated with historical weather, astronomical, and semantic segmentation data to enable analysis of localization performance under different conditions.
The authors benchmark several retrieval-based localization methods, both image-based and LiDAR-based, on the Pit30M dataset. They show that strong convolutional backbones with simple pooling schemes can match the state-of-the-art performance, highlighting the importance of large-scale, diverse datasets like Pit30M for advancing global localization in self-driving vehicles.
The analysis of the results using the dataset's metadata provides insights into the failure modes and complementarity of image and LiDAR-based localization, pointing to future research directions in multi-sensor fusion.
סטטיסטיקה
GPS error is correlated with both image and LiDAR localization error.
Image localization error increases as the sun angle gets closer to the horizon.
LiDAR localization error spikes when 15-20% of points are assigned to dynamic objects.
ציטוטים
"Pit30M is, to the best of our knowledge, the largest benchmark for large-scale localization to date both in terms of images, LiDAR readings, and accurate ground truth information."
"Our dataset includes over 25 000 km and 1 500 hours of driving, resulting in a benchmark that is one to two orders of magnitude larger than those used in previous work."