This is the best validation-selected checkpoint from SafeSynth's controlled four-arm RT-DETRv2-R18 experiment. The selected arm is deliberately real_only: it used all 3,500 frozen real Train images and zero synthetic images. The negative selection result is part of the release, not something hidden by publishing a synthetic arm.

Source and reproducibility: github.com/kuotunyu/SafeSynth
Synthetic ablation dataset: steven0226/safesynth-hard-hat
Base checkpoint: PekingU/rtdetr_v2_r18vd

Model details

Item	Value
Architecture	RT-DETRv2 with ResNet-18 backbone
Classes	`helmet`, `head`, `person`
Training composition	3,500 real Train, 0 synthetic
Validation set	756 real images
Seed	1337
Optimizer steps	10,900
Selected checkpoint	step 1,752 (best frozen-validation checkpoint)
Input preprocessing	Resize to 640x640, rescale by 1/255, no ImageNet normalization
Weight format	Safetensors

The model predicts a helmet box around the helmeted head, a head box for a bare head, and the inherited person class. The third class is unreliable because the source dataset's person annotations are substantially incomplete.

Four-arm result

All arms used the same real Train set, frozen real Validation/Test sets, seed, and 10,900-step optimizer budget. Synthetic arms used 3,500 generated images; their real-image exposure was about half that of the real-only arms.

Arm	Primary AP_small	Primary mAP	Bare-head recall	Real-image exposures
real_only	0.4511	0.5341	0.9875	49.83
standard_aug	0.4236	0.4958	0.9875	49.83
unfiltered_syn	0.3759	0.4597	0.9898	24.91
filtered_syn	0.3664	0.4858	0.9886	24.91

The 1,000-resample, image-level bootstrap intervals for primary AP_small were:

Arm	AP_small (95% CI)
real_only	0.4511 [0.4307, 0.4753]
standard_aug	0.4236 [0.3993, 0.4530]
unfiltered_syn	0.3759 [0.3474, 0.4064]
filtered_syn	0.3664 [0.3426, 0.3956]

AP_small is the mean of helmet and bare-head AP for COCO-small objects, computed in each image's native original annotation coordinates (415/416 pixel edges), not after resizing to 640x640. person is excluded from the primary mAP because its source annotations are known to be poor.

Usage

python

import torch
from PIL import Image
from transformers import RTDetrImageProcessor, RTDetrV2ForObjectDetection

repo_id = "steven0226/safesynth-rtdetrv2-r18"
device = "cuda" if torch.cuda.is_available() else "cpu"

processor = RTDetrImageProcessor.from_pretrained(repo_id)
model = RTDetrV2ForObjectDetection.from_pretrained(repo_id).to(device).eval()
image = Image.open("construction_site.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt").to(device)

with torch.inference_mode():
    outputs = model(**inputs)

target_sizes = torch.tensor([(image.height, image.width)], device=device)
result = processor.post_process_object_detection(
    outputs,
    target_sizes=target_sizes,
    threshold=0.07,
)[0]

for score, label_id, box in zip(
    result["scores"], result["labels"], result["boxes"], strict=True
):
    label = model.config.id2label[int(label_id)]
    print(label, float(score), [round(x, 1) for x in box.tolist()])

The 0.07 threshold was selected on the frozen Validation set for the repository's deployment analysis. Recalibrate it for a new camera, site, class balance, and false-alarm cost. Never tune it on the Test set.

Intended use

This checkpoint is intended for reproducible research, teaching, controlled comparison, and prototyping of hard-hat/bare-head detection. It is not a certified safety product. A human safety process must remain authoritative.

Limitations and responsible use

Do not treat the reported AP as an absolute quality guarantee. SHEL5K re-annotated the same 5,000 source images with 75,570 labels versus 25,502 in the original; all repository claims are relative comparisons on one frozen Test set.
person AP is not a reliable measure because the upstream person labels are sparse and inconsistent.
Training and bootstrap evaluation used one seed. This checkpoint is the validation-selected model for that protocol, not evidence of universal superiority.
The pre-registered synthetic-artifact gate failed (AUC 0.9053 versus a maximum of 0.60). Synthetic augmentation did not robustly improve RT-DETRv2 and showed architecture-sensitive, inconclusive behavior in an RF-DETR cross-check.
Data are drawn from a public construction-worker dataset and may not represent every geography, PPE design, skin tone, workplace, lighting condition, camera, or occlusion pattern.
Do not use the model as the sole basis for discipline, access denial, surveillance, or any decision that can harm a worker.

License

The weights are a fine-tune of PekingU/rtdetr_v2_r18vd, whose Hugging Face repository declares Apache-2.0. These derived weights are released under Apache-2.0. SafeSynth source code is MIT; the immediate training dataset is CC0 1.0. See the linked repositories for the full provenance and notices.