RF-DETR Nano Finetuned on GC10-DET

Fine-tuned RF-DETR Nano 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.

GC10-DET Detection Demo

Usage

Install Dependencies

pip install rfdetr huggingface_hub

Load Model from Hugging Face

from huggingface_hub import hf_hub_download
import rfdetr

weights = hf_hub_download(
    repo_id="dronefreak/gc10det-rfdetr-nano",
    filename="checkpoint_best_total.pth"
)

model = rfdetr.RFDETRNano(pretrain_weights=weights)

Run Inference

detections = model.predict("image.jpg", threshold=0.25)

Performance

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

Metric	Score (%)
mAP@50	70.17
mAP@50-95	38.06
Precision	77.08
Recall	71.04
F1 Score	73.94
Parameters	30.5M
FLOPs	N/A (not published upstream)

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	17.32	10.9
crescent_gap	92.39	61.81
inclusion	50.13	11.4
oil_spot	55.59	22.62
punching_hole	90.98	47.98
rolled_pit	50.0	30.0
silk_spot	66.84	31.89
waist_folding	92.57	62.19
water_spot	86.6	54.29
welding_line	99.31	47.52

This model was evaluated with Supervision's detection metrics, which report mAP/Precision/Recall directly but don't produce a confusion-matrix plot the way Ultralytics' validator does.

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	RF-DETR
Training Toolkit	DetectionBench
Epochs (configured max)	180
Epochs (actually trained)	75
Early Stopping Patience	25
Batch Size	7
Resolution	384
Optimizer	adamw
Learning Rate	0.0001
Seed	42

Repository Contents

checkpoint_best_total.pth
metrics.csv
config.json
gc10det_rfdetr-nano_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}
}

@inproceedings{robinson2026rfdetr,
  title     = {RF-DETR: Real-Time Detection Transformer},
  author    = {Robinson, Isaac and Robicheaux, Peter and Popov, Matvei and Ramanan, Deva and Peri, Neehar},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2511.09554}
}

@article{oquab2023dinov2,
  title={DINOv2: Learning Robust Visual Features without Supervision},
  author={Oquab, Maxime and Darcet, Timoth{\'e}e and Moutakanni, Theo and Vo, Huy and Szafraniec, Marc and Khalidov, Vasil and Fernandez, Pierre and Haziza, Daniel and Massa, Francisco and El-Nouby, Alaaeldin and others},
  journal={arXiv preprint arXiv:2304.07193},
  year={2023}
}