dronefreak/gc10det-yolov8m

YOLOv8m Finetuned on GC10-DET

Fine-tuned YOLOv8m object detector on the GC10-DET benchmark dataset, trained and evaluated as part of DetectionBench -- a framework for reproducibly benchmarking modern object detectors with identical training recipes and evaluation metrics across multiple real-world datasets.

---

Usage

Install Dependencies

pip install ultralytics huggingface_hub

Load Model from Hugging Face

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download(
    repo_id="dronefreak/gc10det-yolov8m",
    filename="best.pt"
)

model = YOLO(weights)

Run Inference

results = model.predict(
    source="image.jpg",
    conf=0.25
)

results[0].show()

---

Performance

Evaluated on the GC10-DET test split, using DetectionBench's standard evaluation pipeline (detectionbench-evaluate).

Metric	Score (%)
mAP@50	71.88
mAP@50-95	38.8
Precision	69.77
Recall	70.84
F1 Score	70.3
Parameters	25.9M
FLOPs	78.9B (at 640 px)

---

GC10-DET Model Zoo

Every model DetectionBench has trained and evaluated on GC10-DET so far, for full transparency -- see DetectionBench for the smaller, curated comparison set used on the project README.

Model	mAP@50	mAP@50-95	Precision	Recall
RF-DETR Small	76.07	42.51	87.86	65.03
RF-DETR Medium	75.93	41.93	78.25	67.76
YOLO26s	75.77	38.15	77.16	74.07
YOLO26n	74.25	38.31	80.5	67.99
YOLO26m	73.97	36.94	75.7	68.19
YOLOv8n	73.25	38.74	67.99	70.87
YOLO11s	72.54	35.07	72.39	66.8
YOLOv8s	72.54	37.79	78.54	65.23
YOLOv8m	71.88	38.8	69.77	70.84
YOLO11n	70.44	40.09	78.93	62.64
RF-DETR Nano	70.17	38.06	77.08	71.04

---

Per-Class Performance

Class	mAP@50	mAP@50-95
crease	30.27	14.18
crescent_gap	91.7	55.7
inclusion	28.56	9.52
oil_spot	47.6	20.49
punching_hole	89.52	50.32
rolled_pit	99.5	79.6
silk_spot	58.12	21.88
waist_folding	94.5	50.04
water_spot	96.01	56.75
welding_line	83.0	29.49

---

Dataset

This model was trained on GC10-DET. For the full dataset description, provenance, license, and citation, see the dataset card:

https://huggingface.co/datasets/dronefreak/GC10-DET

Classes

• crease
• crescent_gap
• inclusion
• oil_spot
• punching_hole
• rolled_pit
• silk_spot
• waist_folding
• water_spot
• welding_line

---

Training Configuration

Setting	Value
Dataset	GC10-DET
Framework	Ultralytics YOLO
Training Toolkit	DetectionBench
Epochs (configured max)	180
Epochs (actually trained)	104
Early Stopping Patience	25
Batch Size	auto (Ultralytics AutoBatch)
Image Size	640
Optimizer	AdamW
Initial Learning Rate	0.001
Seed	0

---

Repository Contents

best.pt
results.csv
args.yaml
BoxPR_curve.png
BoxF1_curve.png
BoxP_curve.png
BoxR_curve.png
confusion_matrix.png
confusion_matrix_normalized.png
val_batch0_pred.jpg
gc10det_yolov8m_showcase.jpg
README.md

---

Related Resources

• GC10-DET dataset card on Hugging Face
• DetectionBench -- reproducible benchmarks for modern object detectors on real-world datasets
• GC10-DET paper (Sensors 2020, doi:10.3390/s20061562)
• GC10-DET GitHub (official data source)

---

Training Framework

This model was trained using DetectionBench, an open-source framework for benchmarking object detectors across multiple real-world datasets with a common pipeline.

Features include:

• A dataset-adapter registry for converting real-world datasets into a canonical format
• Identical training/evaluation recipes across model families (Ultralytics YOLO/RT-DETR, RF-DETR)
• Hardware profiling (latency, FPS, VRAM, parameters, FLOPs)
• One-command reproducibility via versioned Hydra configs

If you find this model useful, please consider starring the repository.

---

Known Limitations

• No official split: GC10-DET's paper defines no train/valid/test division, so this adapter creates a deterministic seeded 80/10/10 split over the sorted-then-shuffled image list -- results are not directly comparable to a paper that uses a different split.
• Small dataset: only 1,840 training images (2,300 total) across 10 classes, so absolute scores are more sensitive to the specific split than on the project's larger datasets, and per-class scores on the rarest classes are noisy.
• Severe class imbalance: silk_spot is 24.8% of all boxes, while crease has only 74 instances (2.1%) across the whole dataset -- its per-class score is measured on very few examples and should be read with caution.
• Sparse, mostly single-defect images: 1.55 instances/image on average (median 1, max 11), unlike the crowded-scene datasets (PKLot/VisDrone/UAVDT) -- this is a localization task on large, easy-to-see boxes (median 3.48% of image area) rather than a small-object or dense-detection problem.
• Different visual domain: GC10-DET is grayscale industrial line-scan imagery of rolled steel surfaces, not a natural-scene photo -- the first industrial-inspection dataset in DetectionBench, so these results say nothing about how these checkpoints would perform on outdoor/natural-scene detection or vice versa.

---

Citation

If you use this model in your research, please consider citing the dataset and the model architecture:

@article{lv2020deep,
  title = {Deep Metallic Surface Defect Detection: The New Benchmark and Detection Network},
  author = {Lv, Xiaoming and Duan, Fajie and Jiang, Jia-jia and Fu, Xiao and Gan, Lin},
  journal = {Sensors},
  volume = {20},
  number = {6},
  pages = {1562},
  year = {2020},
  publisher = {MDPI},
  doi = {10.3390/s20061562}
}

No official YOLOv8 research paper has been published by Ultralytics; this is their own recommended software citation instead:

@software{jocher2023yolov8,
  author = {Glenn Jocher and Ayush Chaurasia and Jing Qiu},
  title = {Ultralytics YOLOv8},
  version = {8.0.0},
  year = {2023},
  url = {https://github.com/ultralytics/ultralytics},
  license = {AGPL-3.0}
}