|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- en |
|
|
base_model: |
|
|
- Ultralytics/YOLOv8 |
|
|
pipeline_tag: object-detection |
|
|
tags: |
|
|
- yolov8 |
|
|
- firearm-detection |
|
|
- object-detection |
|
|
- computer-vision |
|
|
new_version: Subh775/Threat-Detection-RF-DETR |
|
|
--- |
|
|
|
|
|
# Firearm Detection YOLOv8n |
|
|
|
|
|
[](https://opensource.org/licenses/Apache-2.0) |
|
|
[](https://github.com/ultralytics/ultralytics) |
|
|
[](https://ultralytics.com/) |
|
|
[](https://www.python.org/) |
|
|
|
|
|
Try out the more accurate version at: [](https://huggingface.co/Subh775/Threat-Detection-RF-DETR) |
|
|
|
|
|
A high-performance **YOLOv8-nano** model specifically trained for firearm detection in images and videos. This model achieves exceptional accuracy with 89.0% mAP@0.5 and is optimized for real-time inference applications in security and surveillance systems. |
|
|
|
|
|
## Model Overview |
|
|
|
|
|
This model is trained on a comprehensive firearm detection (A custom dataset) containing over 7k high-quality images. The model can detect various types of firearms including pistols, rifles, shotguns, and other weapon types unified under a single "Gun" class for detection of fire-armed weapons |
|
|
|
|
|
**Key Features:** |
|
|
- Single-class detection (Gun) for simplified integration. |
|
|
- High accuracy: 89.0% mAP@0.5, 60.2% mAP@0.5-0.95. |
|
|
- Optimized for real-time inference. |
|
|
- Comprehensive training on diverse firearm types. |
|
|
- Robust performance across different lighting and background conditions. |
|
|
|
|
|
## Sample Demo |
|
|
|
|
|
Here is an example output by the model: |
|
|
|
|
|
<video muted autoplay loop controls src="https://cdn-uploads.huggingface.co/production/uploads/66c6048d0bf40704e4159a23/N6B80__zusImbOoplh_6R.mp4" width=800></video> |
|
|
|
|
|
## Performance Metrics |
|
|
|
|
|
The model demonstrates exceptional performance on the validation dataset after 100 epochs of training: |
|
|
|
|
|
| Metric | Value | |
|
|
|--------|-------| |
|
|
| **mAP@0.5** | **0.890** | |
|
|
| **mAP@0.5-0.95** | **0.602** | |
|
|
| **Precision** | **0.864** | |
|
|
| **Recall** | **0.824** | |
|
|
| **F1-Score** | **0.84** | |
|
|
|
|
|
## Training Results |
|
|
|
|
|
### Training and Validation Curves |
|
|
|
|
|
The training progression over 100 epochs shows consistent improvement across all metrics: |
|
|
|
|
|
 |
|
|
|
|
|
The graphs demonstrate: |
|
|
- **Loss Reduction**: Steady decrease in box_loss, cls_loss, and dfl_loss |
|
|
- **Metric Improvement**: Consistent increase in precision, recall, mAP50, and mAP50-95 |
|
|
- **Convergence**: Stable performance indicating optimal training completion |
|
|
|
|
|
### Dataset Distribution |
|
|
|
|
|
The dataset contains 6,800 gun instances across 7,068 training images, with balanced spatial distribution and varied object sizes for robust detection capabilities. |
|
|
|
|
|
## Model Performance Analysis |
|
|
|
|
|
### Confusion Matrix |
|
|
|
|
|
**Absolute Counts:** |
|
|
 |
|
|
|
|
|
**Normalized Values:** |
|
|
 |
|
|
|
|
|
The confusion matrices show: |
|
|
- **True Positives**: 1,537 correctly identified guns (86% accuracy) |
|
|
- **False Negatives**: 249 missed detections (14% miss rate) |
|
|
- **False Positives**: 324 background misclassifications |
|
|
- **Strong Performance**: High precision with minimal false positive rate |
|
|
|
|
|
### Performance Curves |
|
|
|
|
|
<table align="center"> |
|
|
<tr> |
|
|
<td align="center"> |
|
|
<b>Precision-Recall Curve</b><br> |
|
|
<img src="BoxPR_curve.png" width="400"> |
|
|
</td> |
|
|
<td align="center"> |
|
|
<b>Precision-Confidence Curve</b><br> |
|
|
<img src="BoxP_curve.png" width="400"> |
|
|
</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td align="center"> |
|
|
<b>Recall-Confidence Curve</b><br> |
|
|
<img src="BoxR_curve.png" width="400"> |
|
|
</td> |
|
|
<td align="center"> |
|
|
<b>F1-Confidence Curve</b><br> |
|
|
<img src="BoxF1_curve.png" width="400"> |
|
|
</td> |
|
|
</tr> |
|
|
</table> |
|
|
|
|
|
These curves demonstrate optimal performance at confidence threshold 0.4, balancing precision and recall for practical deployment. |
|
|
|
|
|
## Dataset Information |
|
|
|
|
|
**Training Dataset Composition:** |
|
|
- **Total Images**: 7,068 (after quality filtering) |
|
|
- **Training Set**: 5,642 images |
|
|
- **Validation Set**: 1,426 images |
|
|
- **Total Instances**: 6,800 gun annotations |
|
|
- **Split Ratio**: 80/20 train/validation |
|
|
|
|
|
**Source Datasets:** |
|
|
- Roboflow: 3,822 images (rifles, shotguns, heavy weapons) |
|
|
- Kaggle: 1,946 images (mixed firearms including pistols) |
|
|
- Additional curated sources: 1,444 images |
|
|
|
|
|
**Quality Assurance:** |
|
|
- Duplicate removal: 711 images filtered |
|
|
- Low-quality filtering: 434 images removed |
|
|
- Manual verification and annotation correction |
|
|
|
|
|
--- |
|
|
|
|
|
## Inferencing instructions |
|
|
|
|
|
```bash |
|
|
pip install ultralytics huggingface_hub |
|
|
``` |
|
|
|
|
|
### Basic Inference |
|
|
|
|
|
```python |
|
|
from ultralytics import YOLO |
|
|
from huggingface_hub import hf_hub_download |
|
|
|
|
|
# Download model from Hugging Face Hub |
|
|
model_path = hf_hub_download( |
|
|
repo_id="Subh775/Firearm_Detection_Yolov8n", |
|
|
filename="weights/best.pt" |
|
|
) |
|
|
|
|
|
# Load model |
|
|
model = YOLO(model_path) |
|
|
|
|
|
# Run inference |
|
|
results = model("path/to/your/image.jpg") |
|
|
|
|
|
# Display results |
|
|
for box in results[0].boxes: |
|
|
class_name = model.names[int(box.cls[0])] |
|
|
confidence = box.conf[0] |
|
|
print(f"Detected: {class_name} (Confidence: {confidence:.3f})") |
|
|
|
|
|
# Show annotated image |
|
|
results[0].show() |
|
|
``` |
|
|
|
|
|
### Batch Processing for Videos |
|
|
|
|
|
```python |
|
|
import cv2 |
|
|
from ultralytics import YOLO |
|
|
from huggingface_hub import hf_hub_download |
|
|
import torch |
|
|
from tqdm import tqdm |
|
|
|
|
|
# Configuration |
|
|
MODEL_REPO = "Subh775/Firearm_Detection_Yolov8n" |
|
|
INPUT_VIDEO = "input_video.mp4" |
|
|
OUTPUT_VIDEO = "output_video.mp4" |
|
|
CONFIDENCE_THRESHOLD = 0.4 |
|
|
BATCH_SIZE = 32 # Adjust based on GPU memory |
|
|
|
|
|
# Setup device |
|
|
device = 0 if torch.cuda.is_available() else "cpu" |
|
|
print(f"Using device: {'GPU' if device == 0 else 'CPU'}") |
|
|
|
|
|
# Load model |
|
|
model_path = hf_hub_download(repo_id=MODEL_REPO, filename="weights/best.pt") |
|
|
model = YOLO(model_path) |
|
|
|
|
|
# Process video |
|
|
cap = cv2.VideoCapture(INPUT_VIDEO) |
|
|
frame_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)) |
|
|
frame_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) |
|
|
fps = int(cap.get(cv2.CAP_PROP_FPS)) |
|
|
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) |
|
|
|
|
|
fourcc = cv2.VideoWriter_fourcc(*'mp4v') |
|
|
out = cv2.VideoWriter(OUTPUT_VIDEO, fourcc, fps, (frame_width, frame_height)) |
|
|
|
|
|
frames_batch = [] |
|
|
with tqdm(total=total_frames, desc="Processing video") as pbar: |
|
|
while cap.isOpened(): |
|
|
success, frame = cap.read() |
|
|
if success: |
|
|
frames_batch.append(frame) |
|
|
|
|
|
if len(frames_batch) == BATCH_SIZE: |
|
|
# Batch inference |
|
|
results = model(frames_batch, conf=CONFIDENCE_THRESHOLD, |
|
|
device=device, verbose=False) |
|
|
|
|
|
# Write annotated frames |
|
|
for result in results: |
|
|
annotated_frame = result.plot() |
|
|
out.write(annotated_frame) |
|
|
|
|
|
pbar.update(len(frames_batch)) |
|
|
frames_batch = [] |
|
|
else: |
|
|
break |
|
|
|
|
|
# Process remaining frames |
|
|
if frames_batch: |
|
|
results = model(frames_batch, conf=CONFIDENCE_THRESHOLD, |
|
|
device=device, verbose=False) |
|
|
for result in results: |
|
|
annotated_frame = result.plot() |
|
|
out.write(annotated_frame) |
|
|
pbar.update(len(frames_batch)) |
|
|
|
|
|
cap.release() |
|
|
out.release() |
|
|
print(f"Processed video saved to: {OUTPUT_VIDEO}") |
|
|
``` |
|
|
|
|
|
## Model Architecture |
|
|
|
|
|
- **Base Model**: YOLOv8-nano (yolov8n.pt) |
|
|
- **Input Size**: 640x640 pixels |
|
|
- **Classes**: 1 (Gun) |
|
|
- **Parameters**: ~3.2M |
|
|
- **Model Size**: ~6.2MB |
|
|
- **Inference Speed**: ~4ms per image (GPU) |
|
|
|
|
|
## Training Configuration |
|
|
|
|
|
- **Epochs**: 100 |
|
|
- **Batch Size**: 16 |
|
|
- **Image Size**: 640x640 |
|
|
- **Optimizer**: AdamW |
|
|
- **Learning Rate**: 0.01 (initial) |
|
|
|
|
|
## Limitations and Considerations |
|
|
|
|
|
**Technical Limitations:** |
|
|
- Performance may vary with image quality, lighting, and occlusion |
|
|
- Optimized for common firearm types; may require retraining for specialized weapons |
|
|
- False positives possible with objects resembling firearms (toys, tools) |
|
|
|
|
|
**Recommended Usage:** |
|
|
- Use appropriate confidence thresholds based on application requirements |
|
|
- Implement human review processes for critical decisions |
|
|
- Regular model evaluation and updates for optimal performance |
|
|
- Proper training for operators using the system |
|
|
|
|
|
## License |
|
|
|
|
|
This model is released under the Apache 2.0 License. See the LICENSE file for details. |
|
|
|
|
|
**Disclaimer**: This model is provided for research purposes only. The predictions can not be used to solve real world problems. |
