Faster R-CNN with RoPE-ViT Backbone for Object Detection
This model is a Faster R-CNN object detection model with a RoPE-ViT (Vision Transformer with Rotary Position Embeddings) backbone, trained on the COCO dataset.
Model Description
- Architecture: Faster R-CNN
- Backbone: RoPE-ViT Tiny
- Dataset: COCO
- Task: Object Detection
- Framework: MMDetection
Training Results
| Metric | Value |
|---|---|
| bbox_mAP | 0.0680 |
| bbox_mAP_50 | 0.1510 |
| bbox_mAP_75 | 0.0530 |
| bbox_mAP_s (small) | 0.0360 |
| bbox_mAP_m (medium) | 0.1260 |
| bbox_mAP_l (large) | 0.0640 |
Usage
from mmdet.apis import init_detector, inference_detector
config_file = 'faster_rcnn_rope_vit_tiny_coco.py'
checkpoint_file = 'best_coco_bbox_mAP_epoch_12.pth'
# Initialize the model
model = init_detector(config_file, checkpoint_file, device='cuda:0')
# Inference on an image
result = inference_detector(model, 'demo.jpg')
Training Configuration
The model was trained with the following configuration:
- Input size: 512x512
- Training epochs: 12
- Optimizer: SGD with momentum
- Learning rate scheduler: Step decay
Citation
If you use this model, please cite:
@misc{rope-vit-detection,
author = {VLG IITR},
title = {Faster R-CNN with RoPE-ViT for Object Detection},
year = {2026},
publisher = {Hugging Face},
}
License
This model is released under the Apache 2.0 license.