rtdetr-v2-r50vd-hardhat

This model is a fine-tuned version of PekingU/rtdetr_v2_r50vd on the anindya64/hardhat dataset. It achieves the following results on the evaluation set:

Loss: 7.5564
Map: 0.5656
Map 50: 0.8576
Map 75: 0.6654
Map Small: 0.4886
Map Medium: 0.6765
Map Large: 0.5595
Mar 1: 0.1665
Mar 10: 0.6632
Mar 100: 0.7293
Mar Small: 0.6628
Mar Medium: 0.7915
Mar Large: 0.8442
Map Head: 0.6289
Mar 100 Head: 0.729
Map Helmet: 0.5023
Mar 100 Helmet: 0.7295

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 5e-05
train_batch_size: 8
eval_batch_size: 8
seed: 42
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 0.05
num_epochs: 30

Training results

Training Loss	Epoch	Step	Validation Loss	Map	Map 50	Map 75	Map Small	Map Medium	Map Large	Mar 1	Mar 10	Mar 100	Mar Small	Mar Medium	Mar Large	Map Head	Mar 100 Head	Map Helmet	Mar 100 Helmet
14.4063	1.0	596	7.9687	0.483	0.7777	0.5481	0.428	0.6069	0.4864	0.1448	0.6261	0.7241	0.6689	0.7719	0.8267	0.6099	0.7331	0.3561	0.7151
12.2301	2.0	1192	7.0899	0.6004	0.9325	0.6956	0.5376	0.6876	0.5742	0.1777	0.6645	0.725	0.6737	0.7711	0.7741	0.639	0.7275	0.5619	0.7225
12.0540	3.0	1788	7.1099	0.587	0.9107	0.6936	0.5392	0.6715	0.6388	0.1722	0.6646	0.7303	0.6745	0.7805	0.8468	0.6396	0.7305	0.5344	0.7301
11.9553	4.0	2384	7.1824	0.6138	0.9389	0.7173	0.5527	0.6953	0.7044	0.1763	0.6707	0.7339	0.6715	0.7892	0.8503	0.6632	0.7446	0.5643	0.7232
11.4423	5.0	2980	7.0358	0.6108	0.9207	0.7263	0.5615	0.7015	0.629	0.1768	0.68	0.7404	0.6825	0.794	0.8359	0.6646	0.7495	0.5571	0.7314
11.9962	6.0	3576	7.0091	0.6247	0.9419	0.7349	0.5552	0.7113	0.7155	0.1812	0.6836	0.7412	0.6809	0.7965	0.8341	0.6594	0.7403	0.59	0.742
11.4962	7.0	4172	7.1070	0.6341	0.9486	0.7524	0.5638	0.7169	0.7335	0.1813	0.6851	0.7403	0.6786	0.7979	0.8241	0.6638	0.7459	0.6044	0.7346
11.3964	8.0	4768	7.1871	0.6166	0.9311	0.7269	0.5425	0.7032	0.7206	0.1784	0.683	0.7384	0.6729	0.7977	0.8362	0.663	0.738	0.5702	0.7388
11.0877	9.0	5364	7.1819	0.6186	0.9335	0.7344	0.5561	0.7052	0.684	0.1782	0.6819	0.7333	0.6719	0.7888	0.8446	0.6584	0.7381	0.5789	0.7284
11.0089	10.0	5960	7.2926	0.5958	0.901	0.7036	0.5254	0.6988	0.5998	0.1744	0.6779	0.734	0.6742	0.7894	0.8104	0.6589	0.7324	0.5327	0.7355
10.9968	11.0	6556	7.1043	0.6313	0.9489	0.7617	0.5579	0.7107	0.7015	0.18	0.6807	0.7404	0.681	0.794	0.8323	0.6558	0.7424	0.6069	0.7384
10.9626	12.0	7152	7.3803	0.5928	0.8888	0.6999	0.5215	0.6999	0.5475	0.1711	0.6799	0.7462	0.6847	0.802	0.8191	0.6476	0.7529	0.5379	0.7396
11.1005	13.0	7748	7.2846	0.6189	0.9248	0.7431	0.5434	0.7099	0.6319	0.1795	0.6816	0.7412	0.6792	0.7983	0.8003	0.6546	0.7456	0.5832	0.7368
10.9196	14.0	8344	7.3971	0.5822	0.8707	0.6941	0.5134	0.6934	0.5429	0.1705	0.6749	0.7437	0.6826	0.8008	0.8472	0.6356	0.7437	0.5288	0.7437
10.5408	15.0	8940	7.3988	0.5987	0.8976	0.7074	0.5337	0.6961	0.5445	0.1748	0.675	0.7416	0.6779	0.8002	0.8433	0.6575	0.7517	0.5399	0.7316
10.8451	16.0	9536	7.5564	0.5656	0.8576	0.6654	0.4886	0.6765	0.5595	0.1665	0.6632	0.7293	0.6628	0.7915	0.8442	0.6289	0.729	0.5023	0.7295

Framework versions

Transformers 5.17.0
Pytorch 2.14.0+cu130
Datasets 5.0.1
Tokenizers 0.23.2

Как пользоваться

from transformers import pipeline

detector = pipeline("object-detection", model="KuBaN1900/rtdetr-v2-r50vd-hardhat")
results = detector("photo.jpg", threshold=0.05)

Рекомендуемый порог уверенности — 0.05: он подобран по максимуму F1 на валидационной выборке. Если цена пропуска выше цены ложной тревоги (например, контроль касок на объекте), снижайте порог до 0.05 — это точка максимума F2.

Качество на тестовой выборке

Метрика	Значение
mAP@0.5:0.95	0.601
mAP@0.5	0.901
mAP@0.75	0.718
mAP small	0.530
mAP medium	0.682
mAP large	0.736
mAR@100	0.748

По классам:

класс	mAP@0.5:0.95	mAR@100	объектов в тесте
head	0.655	0.752	1803
helmet	0.546	0.744	4863

Как обучалась

Базовая модель: PekingU/rtdetr_v2_r50vd
Датасет: anindya64/hardhat, сплит train разделён на train/val как 90/10
Вход: 640x640, аугментации albumentations (flip, affine, яркость/контраст, HSV, motion blur)
Оптимизатор: AdamW, LR 5e-05 (бэкбон: x0.1), cosine schedule, warmup 0.05
Лучший чекпоинт выбран по mAP@0.5 на валидации, логи эксперимента — в TensorBoard (rtdetr-v2-r50vd-hardhat/runs)