RF-DETR Plus
https://github.com/roboflow/rf-detr_plus- Category
- AI Tools
- Rank
- No. 902Tools index
Previous survey · No. 910 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- roboflow
- GitHub
- 57 stars
- Latest release
- 1.0.2
- Date
About
Extension of RF-DETR unlocking larger, more powerful detection models for highest-accuracy computer vision tasks.
What it does
RF-DETR Plus supplies two object-detection model sizes that plug into the training and inference stack of the core RF-DETR package. Both use a DINOv2 vision transformer with larger feature dimensions and high input resolutions. Users can predict from images or fine-tune on their own datasets.
Why it's ranked here
The case rests on measurable gains within one model family. The larger variants lead the supplied COCO and RF100-VL benchmark results, while documented TensorRT tests show batch-one latency of 11.5 and 17.2 milliseconds on an NVIDIA T4. That makes the package compelling when detection accuracy outweighs model size and licensing flexibility.
What's good
It preserves the core package’s training and inference workflow, so existing users gain larger models without adopting a separate framework. Published benchmarks state hardware, precision, batch size, resolution, parameter count, accuracy, and latency. Weight records also include integrity hashes, and the test suite exercises both basic prediction and COCO evaluation on CPU and GPU configurations.
Tradeoffs
The license is the largest constraint. Use requires a platform agreement and an account in good standing, while reverse engineering and bypassing usage tracking are prohibited. The models contain roughly 126 million parameters, far above the core lineup, and use 700 or 880 pixel inputs. The package is marked Alpha and declares Unix, POSIX, and macOS support, but not Windows.
How to use it well
Choose it when an existing RF-DETR workflow needs the family’s strongest documented detection results and can accept the platform terms. Benchmark both sizes on your target hardware before selecting one, since their latency and resolution differ. Keep the core package for smaller models and broader Apache-licensed use. This extension does not replace the underlying training, inference, export, or deployment stack.
Technical notes+
pyproject.toml defines a setuptools package for Python 3.10 or newer, depends on rfdetr>=1.8.0,<2, includes py.typed, and exposes no CLI entry point. src/rfdetr_plus/models/detection.py subclasses RFDETR with RFDETRXLarge and RFDETR2XLarge, using 700 and 880 resolutions, a 512 hidden dimension, five decoder layers, and DINOv2 windowed-base encoders. Its constructors discard accept_platform_model_license before delegating, so license enforcement is contractual rather than implemented by that argument in this file. src/rfdetr_plus/assets/model_weights.py registers two hosted checkpoints with MD5 hashes. tests/test_inference.py covers prediction plus thresholded COCO mAP and F1 benchmarks, with heavier cases marked for GPU execution.
Observed
- License
- Platform Model License 1.0 applies to code, model weights, checkpoints, artifacts, and documentation.
- Primary language
- Python, with Python 3.10 or newer required.
- Packaging
- Distributed as the rfdetr-plus Python package through pip, with setuptools as the build backend.
- Interface
- Python library exposing two detection model classes; no command-line entry point is declared.
- Core dependency
- Requires rfdetr version 1.8 or newer and below 2.
- Platform support
- Package classifiers declare POSIX, Unix, and macOS support.
- Testing structure
- The repository includes pytest inference tests with CPU and GPU-marked COCO benchmark cases.
Read from README.md, pyproject.toml, src/rfdetr_plus/__init__.py, src/rfdetr_plus/assets/__init__.py, src/rfdetr_plus/models/__init__.py, src/rfdetr_plus/models/detection.py, src/rfdetr_plus/models/downloads.py, src/rfdetr_plus/assets/model_weights.py, tests/conftest.py, tests/test_inference.py, LICENSE, AGENTS.md, CLAUDE.md, .codecov.yml, .pre-commit-config.yaml.
What it can do
Detect objects in images with high accuracy
Digital images or image files → Object detection results with bounding boxes and confidence scores
Process large-scale datasets for object detection
Large collections of images or video frames → Batch detection results with identified objects and locations
Train custom object detection models
Labeled training datasets and model configuration parameters → Trained detection models optimized for specific use cases
Perform real-time object detection
Live video streams or camera feeds → Real-time detection annotations and object tracking data
Generate detection confidence metrics
Images with detected objects → Accuracy scores and confidence levels for each detection
Export detection results in multiple formats
Processed detection data → Results in JSON, XML, CSV, or annotation file formats
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.