harshdadiya-wappnet/phone_detction_rfdetr_base

Model Card for Phone Detection (RF-DETR)

This model performs single-class object detection to identify and localize mobile phones in images.
It is fine-tuned from stevenbucaille/rf-detr-base using the harshdadiya-wappnet/phone_detection dataset.

---

Model Details

Model Description

• Developed by: Harsh Dadiya
• Funded by: Wappnet Systems
• Shared by: harshdadiya-wappnet
• Model type: Object Detection
• Language(s): Not applicable (vision-only)
• License: Apache-2.0
• Finetuned from model: stevenbucaille/rf-detr-base

Model Sources

• Repository: Hugging Face Hub

---

Dataset Details

Dataset Split Statistics

Class Name	Total	Train	Validation	Test
mobile_phone	293	198	68	27

• Number of classes: 1
• Annotation format: COCO
• Task: Bounding box detection

---

Uses

Direct Use

• Detecting mobile phones in still images
• Bounding box localization for computer vision pipelines
• Benchmarking single-class object detection models

Downstream Use

• Further fine-tuning on domain-specific phone imagery
• Integration into monitoring, compliance, or vision-based analysis systems

Out-of-Scope Use

• Multi-class object detection
• Video-based temporal detection
• Classification-only tasks without localization

---

Bias, Risks, and Limitations

• Trained on a single-class dataset, limiting generalization to other object categories
• Performance may degrade on images with extreme occlusion, motion blur, or very small phones
• Dataset size is relatively small and may not cover all real-world variations

---

Training Details

Training Data

• Dataset: harshdadiya-wappnet/phone_detection
• Classes: mobile_phone
• Format: COCO

Training Procedure

• Fine-tuning performed on RF-DETR base model
• Training tracked using Weights & Biases (W&B)

Training Hyperparameters

• Training regime: Not specified

---

Training Metrics and Run History

Run Summary (Final Epoch: 39)

Metric	Value
Loss / Train	2.52222
Loss / Test	4.15493
Base AP@50	0.96172
Base AP@50–90	0.79702
Base AR@50–90	0.85217
EMA AP@50	0.97117
EMA AP@50–90	0.84201
EMA AR@50–90	0.88551

---

Training Curves

The following plots were recorded during fine-tuning:

• Training Loss vs Epoch
• Validation Loss vs Epoch
• AP@50 vs Epoch
• AP@50–90 vs Epoch
• Recall (AR@50–90) vs Epoch

---

Weights & Biases Logs

• Run: phone_detection_via_rfdetr
• Run URL: harshd-wappnet-wappnet-systems-pvt-/uncategorized/g0raad5u

Artifacts:

• Auto-checkpoints saved
• Best model checkpoint saved
• TensorBoard logs generated

---

Evaluation

Metrics

• mAP@50: 83.6%
• Precision: 86.3%
• Recall: 82.1%

Summary

The model demonstrates strong performance for single-class phone detection, with high precision and balanced recall across evaluation splits.

---

Technical Specifications

Model Architecture

• RF-DETR (Transformer-based object detection)
• Bounding box regression + class prediction

---

Model Card Authors

Harsh Dadiya

---