Benchmark & Challenges 2026

CUBIT
BENCHMARK

"A Large-Scale Benchmark for Infrastructure Defect Assessment and Physical Quantification Challenges."

Participate Challenge

12,500+

Images

32,400+

Instances

Download & References

Shared Resources

The linked drive folder contains the training materials, test logs, and verification notes used to support the benchmark release.

Open Google Drive

Training files and notes

Training folders include the reference files plus short notes that explain the setup and file naming.

Test logs

The test logs are collected separately so reviewers can trace the inference outputs and evaluation flow.

Official repository comparison

During the repair process, we keep a comparison with the official repository to verify consistency and avoid regressions.

Top-5 repeated training validation

We also keep repeated-training checks for the Top-5 models to confirm the reported ranking is stable.

Academic Benchmark

Performance Leaderboard

Rank	Model	Crack AP	Spalling AP	Latency
#01	YOLOv6-l	85.7%	91.7%	15.9ms
#02	YOLOv5-x	81.2%	88.4%	28.4ms
#03	YOLOX-x	83.9%	89.5%	41.2ms
#04	Faster R-CNN (ResNet50)	72.5%	71.5%	55.0ms

Participate in CUBIT Challenges

Submit your models to our automated evaluation server. We follow the MS COCO (101 points) and Pascal VOC protocols.

Dataset Taxonomy

Data Modality

UAV RGB Imagery (8K/4K/HD)

Defect Classes

Cracks (Linear, Branch, Web), Spalling, Moisture

Data Split

70% Train, 10% Val, 20% Robust Test

Metric Scale

Pixel-to-MM (Calibrated via GSD)

Physical Severity Grading

LowSI < 0.25

Crack: < 0.2 mm

Action: Routine monitoring

ModerateSI 0.25 - 0.50

Crack: 0.2 - 0.5 mm

Action: Repair scheduling

SevereSI 0.50 - 0.75

Crack: > 0.5 mm

Action: Urgent repair

CriticalSI > 0.75

Crack: Severe structural damage

Action: Immediate structural assessment

* Based on BS ISO 15686-7:2017 and HK Surveyors Practice.

Protocols

Submission & Metrics

CUBIT defines standardized data structures for fair evaluation across all participating models.

Task 1: Detection

Focuses on high-resolution bounding box localization. Each submission must contain category-specific .txt files in the following format:

[imgname] [score] [xmin] [ymin] [xmax] [ymax]
pavement_001 0.985 450 120 890 340
pavement_001 0.742 1200 4500 1350 4800
...

Challenge Rules

• Images must not be resized below 1024x1024 during inference.
• Results are evaluated using Pascal VOC AP@0.5 and COCO AP@0.5:0.95.

Task 2: InSeg & Quant

Focuses on pixel-accurate masks and physical quantification. Submissions should follow the COCO JSON format with an additional metric field.

{
  "image_id": 405,
  "segmentation": [45.1, 89.2, ...],
  "physical_metric": 0.42, // mm for cracks
  "physical_unit": "mm"
}

Quantification Metric

Evaluated by Root Mean Square Error (RMSE) against ground-truth physical measurements.

CUBITBENCHMARK