Railway Track Defect Detection
Synthetic railway track imagery featuring broken rails, sleepers, and trackside obstructions. Diverse environmental conditions and structural anomalies provide robust training data for automated rail inspection systems. Obstruction classes provide support for false positive detection filtering.
Sample Frames
20 annotated samples drawn from the train split. Toggle annotations to inspect bounding-box quality.
Class Distribution
9 annotation classes · 876 total images. Sorted by object count, descending.
Class Balance
Per-class counts, frequency, and average bounding-box area. Sort any column to surface the rarest or most prevalent classes for re-balancing.
| Class | Images↓ with class | Objects total | Per image average | Area % of frame |
|---|---|---|---|---|
| BrokenRail | 847 | 860 | 1.02 | 1.94% |
| Ignore | 494 | 494 | 1 | 0.98% |
| Sleeper | 464 | 920 | 1.98 | 4.45% |
| Vegetation | 250 | 1,007 | 4.03 | 0.66% |
| Droppings | 188 | 218 | 1.16 | 0.88% |
| Shadow | 163 | 569 | 3.49 | 1.90% |
| Cables | 123 | 244 | 1.98 | 0.75% |
| SleeperSupport | 91 | 94 | 1.03 | 0.54% |
| HalfBrokenRail | 20 | 20 | 1 | 1.22% |
Co-occurrence Matrix
How frequently pairs of classes appear in the same image. Diagonal cells show standalone image count for that class. Useful for spotting biased or correlated labels.
| BrokenRail | Vegetation | Sleeper | Cables | Droppings | Ignore | Shadow | SleeperSupport | HalfBrokenRail | |
|---|---|---|---|---|---|---|---|---|---|
| BrokenRail | 847 | 234 | 450 | 117 | 183 | 481 | 158 | 88 | |
| Vegetation | 234 | 250 | 143 | 41 | 44 | 148 | 23 | 11 | |
| Sleeper | 450 | 143 | 464 | 88 | 115 | 359 | 84 | 1 | 11 |
| Cables | 117 | 41 | 88 | 123 | 3 | 93 | 1 | 3 | |
| Droppings | 183 | 44 | 115 | 3 | 188 | 125 | 3 | 26 | 3 |
| Ignore | 481 | 148 | 359 | 93 | 125 | 494 | 100 | 10 | |
| Shadow | 158 | 23 | 84 | 1 | 3 | 100 | 163 | 19 | 2 |
| SleeperSupport | 88 | 1 | 26 | 19 | 91 | 1 | |||
| HalfBrokenRail | 11 | 11 | 3 | 3 | 10 | 2 | 1 | 20 |
Average Object Area
Each rectangle is one class, sized by the average area its bounding boxes occupy as a percentage of the frame. Surfaces tiny vs. dominant objects at a glance.
Spatial Distribution
Where annotations of each class tend to fall across the frame. Brighter regions indicate higher density — useful for detecting positional bias in your training data.
Model Performance
Validation metrics from a YOLOv8 detector trained on this dataset. Reference checkpoint: yolov8m.pt.
Validation curves over training
Dataset Metadata
| Annotation format | YOLO |
|---|---|
| Total images | 876 |
| Classes | 9 |
| Resolution | Mixed |
| Last updated | 2026-05 |
Interested in this dataset?
Get in touch for sample access, custom variants, or licensing terms — typical turnaround is two business days.