https://github.com/bluemirrors/cvu
Computer Vision deployment tools for dummies and experts. CVU aims at making CV pipelines easier to build and consistent around platforms, devices, and models.
https://github.com/bluemirrors/cvu
computer-vision object-detection onnxruntime pytorch tensorflow2 tensorrt yolov5
Last synced: 1 day ago
JSON representation
Computer Vision deployment tools for dummies and experts. CVU aims at making CV pipelines easier to build and consistent around platforms, devices, and models.
- Host: GitHub
- URL: https://github.com/bluemirrors/cvu
- Owner: BlueMirrors
- License: apache-2.0
- Created: 2021-05-15T03:49:05.000Z (almost 4 years ago)
- Default Branch: master
- Last Pushed: 2023-08-20T06:38:15.000Z (over 1 year ago)
- Last Synced: 2025-05-08T21:16:40.359Z (1 day ago)
- Topics: computer-vision, object-detection, onnxruntime, pytorch, tensorflow2, tensorrt, yolov5
- Language: Python
- Homepage: https://pypi.org/project/cvu-python/
- Size: 39 MB
- Stars: 88
- Watchers: 3
- Forks: 19
- Open Issues: 3
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# CVU: Computer Vision Utils 
[](https://www.codefactor.io/repository/github/bluemirrors/cvu) [](https://github.com/mkenney/software-guides/blob/master/STABILITY-BADGES.md#alpha)
[](https://www.python.org/) [](https://pepy.tech/project/cvu-python) [](https://colab.research.google.com/drive/1FvebFw40Bm0bUHWCgS0-iuYp8AKLIfSh?usp=sharing)
Computer Vision pipeline framework with SOTA components for dummies and experts.
Whether you are developing an end-to-end computer vision pipeline or just looking to use some quick computer vision in your project, CVU can help! Designed to be used by both the expert and the novice, CVU aims at making CV pipelines easier to build and consistent around platforms, devices and models.
Code Example 
```bash
pip install cvu-python
```
CVU lets you create end-to-end pipelines with various SOTA/customizable components. With a focus on a common component interface, you naturally create a loosely coupled pipeline with most of the implementation details hidden. Because of this, you can combine any number of CVU components, in any order, to create a pipeline of your need. You can set and switch between one or multiple pipeline input sources (eg. an image, folder, video, or live stream) and output sinks (eg. video file, image with results drawn, TXT/JSON dumps, etc.)
It also comes with optional + customizable default settings which can run a benchmark on your platform/machine to optimally choose dependencies based on accuracy and latency preferences. CVU can also automatically switch/select target devices (CPU, GPU, TPU), computation backends (TF, PyTorch, ONNX, TensorRT, TFLite), and models (small, big, etc.) based on where the pipeline is running.
Currently, CVU only provides Object Detection, but we are in the process to support Segmentation, Background removal, Tracking, and Image text matching out of the box.
# Index 📋
- [Getting Started](#cvu--says-hi) [](https://colab.research.google.com/drive/1FvebFw40Bm0bUHWCgS0-iuYp8AKLIfSh?usp=sharing)
- [Why CVU?](https://github.com/BlueMirrors/cvu/wiki)
- [Object Detection (YOLOv5)](https://github.com/BlueMirrors/cvu/wiki/YOLOv5-object-detection) [](https://colab.research.google.com/drive/1FvebFw40Bm0bUHWCgS0-iuYp8AKLIfSh?usp=sharing)
- [TensorRT](https://github.com/BlueMirrors/cvu/wiki/YOLOv5-TensorRT)
- [Torch](https://github.com/BlueMirrors/cvu/wiki/YOLOv5-Torch)
- [ONNX](https://github.com/BlueMirrors/cvu/wiki/YOLOv5-ONNX)
- [TensorFlow](https://github.com/BlueMirrors/cvu/wiki/YOLOv5-TensorFlow)
- [TFLite](https://github.com/BlueMirrors/cvu/wiki/YOLOv5-TFLite)
- [Devices (CPU, GPU, TPU)](#devices)
- [Benchmark-Tool (YOLOv5)](https://github.com/BlueMirrors/cvu/wiki/Benchmark-tool)
- [Benchmarks Results (YOLOv5)](https://github.com/BlueMirrors/cvu/wiki/YOLOv5-benchmarking)
- [Precission Accuracy (YOLOv5)](https://github.com/BlueMirrors/cvu/wiki/YOLOv5-object-detection#precission-accuracy-yolov5)
- [Examples](https://github.com/BlueMirrors/cvu/tree/master/examples)
- [References](#references)
# CVU Says Hi!
[Index](#index-)
How many installation-steps and lines of code will you need to run object detection on a video with a TensorRT backend? How complicated is it be to test that pipeline in Colab?
With CVU , you just need the following! No extra installation steps needed to run on Colab, just pip install our tool, and you're all set to go!
```python
from vidsz.opencv import Reader, Writer
from cvu.detector import Detector
# set video reader and writer, you can also use normal OpenCV
reader = Reader("example.mp4")
writer = Writer(reader, name="output.mp4")
# create detector with tensorrt backend having fp16 precision by default
detector = Detector(classes="coco", backend="tensorrt")
# process frames
for frame in reader:
# make predictions.
preds = detector(frame)
# draw it on frame
preds.draw(frame)
# write it to output
writer.write(frame)
writer.release()
reader.release()
```
Want to use less lines of code? How about this!
```python
from cvu.detector import Detector
from vidsz.opencv import Reader, Writer
detector = Detector(classes="coco", backend="tensorrt")
with Reader("example.mp4") as reader:
with Writer(reader, name="output.mp4") as writer:
writer.write_all(map(lambda frame:detector(frame).draw(frame), reader))
```
Want to switch to non-cuda device? Just set `device="cpu"`, and backend to `"onnx"`, `"tflite"`, `"torch"` or `"tensorflow"`.
```python
detector = Detector(classes="coco", backend="onnx", device="cpu")
```
Want to use TPU? Just set `device="tpu"` and choose a supported backend (only `"tensorflow"` supported as of the latest release)
```python
detector = Detector(classes="coco", backend="tensorflow", device="tpu")
```
You can change devices, platforms and backends as much as you want, without having to change your pipeline.
# Devices
[Index](#index-)
### Support Info
Following is latest support matrix
| Device | TensorFlow | Torch | TFLite | ONNX | TensorRT |
| ------ | ---------- | ----- | ------ | ---- | -------- |
| GPU | ✅ | ✅ | ❌ | ✅ | ✅ |
| CPU | ✅ | ✅ | ✅ | ✅ | ❌ |
| TPU | ✅ | ❌ | ❌ | ❌ | ❌ |
### Recommended Backends (in order)
Based on FPS performance and various benchmarks
- GPU: `TensorRT` > `Torch` > `ONNX` > `TensorFlow`
- CPU: `ONNX` > `TFLite` > `TensorFlow` > `Torch`
- TPU: `TensorFlow`
# References
- **_Logo-Attribution_**
Designed by roserodionova / Freepik
- [Yolov5 (Default Object Detection Model)](https://github.com/ultralytics/yolov5)