25,000 high-resolution street-level images featuring 124 semantic object categories with dense, pixel-accurate human annotations. The collection includes 70 instance-specific labels for fine-grained object recognition across diverse global environments.
Use Cases
- Train semantic segmentation models to classify pixels into 124 distinct object categories
- Develop instance segmentation architectures using the 70 instance-specific labels for object detection and mask generation
- Evaluate the robustness of autonomous driving perception systems across diverse global street scenes
Strengths
- 25,000 high-resolution images with global geographic coverage
- 124 semantic object categories for dense pixel-level classification
- 70 instance-specific labels for individual object segmentation
- Pixel-accurate human annotations for fine-grained ground truth