Conditional-DETR ResNet-50 Signature Detector
tech4humans/conditional-detr-50-signature-detector
published Jun 2025 · updated Jun 2025
Conditional-DETR ResNet-50 Signature Detector is a detection model that locates handwritten signatures in document images using a Conditional-DETR architecture with a ResNet-50 backbone.
specs
| Task | Object Detection |
| Architecture | Conditional-DETR with ResNet-50 backbone |
| Parameters | Not specified |
| License | Apache 2.0 |
about this model
Conditional-DETR ResNet-50 signature detector is a transformer-based object detection model fine-tuned to locate handwritten signatures in document images. It achieved the highest mean average precision at IoU 0.5 ([email protected]) of 93.65% among 21 object detection architectures evaluated, including YOLOv8, YOLOv10, YOLO11, RT-DETR, DETR, and YOLOS variants. The model's [email protected]:0.95 is 0.653.
Training Dataset
The model was trained on 2,819 document images (640x640 pixels) combining Tobacco800 and signatures-xc8up datasets, split into 70% training, 15% validation, and 15% test sets in COCO JSON format.
Benchmark Summary
The following table compares the top-performing models from the evaluation. The full comparison of 21 detectors is available in the associated article.
| Model | mAP50 | mAP50-95 | CPU Inference (ms) |
|---|---|---|---|
| conditional-detr-resnet-50 | 0.9365 | 0.6533 | 476.8 |
| rtdetr-l | 0.9271 | 0.6224 | 583.6 |
| yolov8m | 0.8753 | 0.6655 | 401.2 |
| yolos-base | 0.9012 | 0.5836 | 1706.5 |
This model is hosted on gigarouter as a managed API with OpenAI-compatible endpoints, enabling integration into document processing pipelines without infrastructure management.
best for
- ·Automated signature detection in scanned documents
- ·Document processing pipelines for legal or financial forms
- ·Real-time signature verification preprocessing
FAQ
The model achieves 93.65% [email protected] on the test set, the highest among all architectures evaluated.
The model expects images resized to 640x640 pixels, typically in COCO JSON format.
Conditional-DETR ResNet-50 has a CPU inference time of ~476.8 ms, slower than YOLOv8s (~216.6 ms) but with higher [email protected].
The model is released under the Apache 2.0 license.
Use the gigarouter OpenAI-compatible endpoint with an API key to send images and receive detection results.
We're benchmarking and onboarding Conditional-DETR ResNet-50 Signature Detector as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.