Home

Awesome

Deep Vision Processing Build

Deep computer-vision algorithms for Processing.

The idea behind this library is to provide a simple way to use (inference) machine learning algorithms for computer vision tasks inside Processing. Mainly portability and easy-to-use are the primary goals of this library. Starting with version 0.6.0 CUDA inferencing support is built into the library (Windows & Linux).

Caution: The API is still in development and can change at any time.

Pose

Lightweight OpenPose Example

Install

It is recommended to use the contribution manager in the Processing app to install the library.

image

Manual

Download the latest prebuilt version from the release sections and install it into your Processing library folder.

Usage

The base of the library is the DeepVision class. It is used to download the pretrained models and create new networks.

import ch.bildspur.vision.*;
import ch.bildspur.vision.network.*;
import ch.bildspur.vision.result.*;

DeepVision vision = new DeepVision(this);

Usually it makes sense to define the network globally for your sketch and create it in setup. The create method downloads the pre-trained weights if they are not already existing. The network first has to be created and then be setup.

YOLONetwork network;

void setup() {
  // create the network & download the pre-trained models
  network = vision.createYOLOv3();

  // load the model
  network.setup();
  
  // set network settings (optional)
  network.setConfidenceThreshold(0.2f);
  
  ...
}

By default, the weights are stored in the library folder of Processing. If you want to download them to the sketch folder, use the following command:

// download to library folder
vision.storeNetworksGlobal();

// download to sketch/networks
vision.storeNetworksInSketch();

Each network has a run() method, which takes an image as a parameter and outputs a result. You can just throw in any PImage and the library starts processing it.

PImage myImg = loadImage("hello.jpg");
ResultList<ObjectDetectionResult> detections = network.run(myImg);

Please have a look at the specific networks for further information or at the examples.

OpenCL Backend Support

With version 0.8.1 by default if OpenCL is enabled, it will be used as backend. If CUDA is enabled too, CUDA will be preferred. It is possible to force the CPU backend by setting the following option:

DeepVision vision = new DeepVision(this);
vision.setUseDefaultBackend(true);

CUDA Backend Support

With version 0.6.0 it is possible to download the CUDA bundled libraries. This enables to run most of the DNN's on CUDA enabled graphics cards. For most networks this is necessary to run them in real-time. If you have the cuda-bundled version installed and run deep-vision on a Linux or Windows with an NVIDIA graphics card, you are able to enable the CUDA backend:

// Second parameter (enableCUDABackend) enables CUDA
DeepVision vision = new DeepVision(this, true);

If the second parameter is unset, the library will check if a CUDA enabled device is available and enables the backend likewise. It is possible to check if CUDA backend has been enabled by the following method:

println("Is CUDA Enabled: " + vision.isCUDABackendEnabled());

If CUDA is enabled but the hardware does not support it, Processing will show you a warning and run the networks on the CPU.

Networks

Here you find a list of implemented networks:

The following list shows the networks that are on the list to be implemented (⚑️ already in progress):

Object Detection

Locating one or multiple predefined objects in an image is the task of the object detection networks.

YOLO

YOLO Example

The result of these networks is usually a list of ObjectDetectionResult.

ObjectDetectionNetwork net = vision.createYOLOv3();
net.setup();

// detect new objects
ResultList<ObjectDetectionResult> detections = net.run(image);

for (ObjectDetectionResult detection : detections) {
    println(detection.getClassName() + "\t[" + detection.getConfidence() + "]");
}

Every object detection result contains the following fields:

YOLO [Paper]

YOLO a very fast and accurate single shot network. The pre-trained model is trained on the 80 classes COCO dataset. There are three different weights & models available in the repository:

// setup the network
YOLONetwork net = vision.createYOLOv4();
YOLONetwork net = vision.createYOLOv4Tiny();
YOLONetwork net = vision.createYOLOv3();
YOLONetwork net = vision.createYOLOv3SPP();
YOLONetwork net = vision.createYOLOv3Tiny();
YOLONetwork net = vision.createYOLOv5n();
YOLONetwork net = vision.createYOLOv5s();
YOLONetwork net = vision.createYOLOv5m();
YOLONetwork net = vision.createYOLOv5l();
YOLONetwork net = vision.createYOLOv5x();

// set confidence threshold
net.setConfidenceThreshold(0.2f);

YOLOv5

Since version 0.9.0 YOLOv5 is implemented as well. It uses the pre-trained models converted into the ONNX format. At the moment YOLOv5 does not work well with the implemented NMS. To adjust the settings of the NMS use the following functions.

// set confidence threshold
net.setConfidenceThreshold(0.2f);

// set confidence threshold
net.set(0.2f);

// set the IoU threshold (overlapping of the bounding boxes)
net.setNmsThreshold(0.4f);

// set how many objects should be taken into account for nms
// 0 means all objects
net.setTopK(100);

SSDMobileNetV2 [Paper]

This network is a single shot detector based on the mobilenetv2 architecture. It is pre-trained on the 90 classes COCO dataset and is really fast.

SSDMobileNetwork net = vision.createMobileNetV2();

Handtracking [Project]

This is a pre-trained SSD MobilenetV2 network to detect hands.

SSDMobileNetwork net = vision.createHandDetector();

TextBoxes [Paper]

TextBoxes is a scene text detector in the wild based on SSD MobileNet. It is able to detect text in a scene and return its location.

TextBoxesNetwork net = vision.createTextBoxesDetector();

Ultra-Light-Fast-Generic-Face-Detector [Project]

ULFG Face Detector is a very fast CNN based face detector which reaches up to 40 FPS on a MacBook Pro. The face detector comes with four different pre-trained weights:

ULFGFaceDetectionNetwork net = vision.createULFGFaceDetectorRFB640();
ULFGFaceDetectionNetwork net = vision.createULFGFaceDetectorRFB320();
ULFGFaceDetectionNetwork net = vision.createULFGFaceDetectorSlim640();
ULFGFaceDetectionNetwork net = vision.createULFGFaceDetectorSlim320();

The detector detects only the frontal face part and not the complete head. Most algorithms that run on results of face detections need a rectangular detection shape.

Cascade Classifier [Paper]

The cascade classifier detector is based on boosting and very common as pre-processor for many classifiers.

CascadeClassifierNetwork net = vision.createCascadeFrontalFace();

Object Recognition

tbd

KeyPoint Detection

tbd

Classification

tbd

Depth Estimation

MidasNet

Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer

MidasNet

Image Processing

tbd

Pipeline

It is possible to create network pipelines to use for example a face-detection network and different classifier for each face. This is not yet documented so you have to check out the test code: HumanAttributesPipelineTest.java#L36-L41

Build

Run gradle to build a new release package under /release/deepvision.zip:

# windows
gradlew.bat releaseProcessingLib

# mac / unix
./gradlew releaseProcessingLib

Cuda Support

To build with CUDA support enable the property cuda:

gradlew.bat releaseProcessingLib -Pcuda -Pdisable-fatjar

This will take several minutes and result in a 5.3 GB folder. disable-fatjar prevents form creating a fatjar, which would be too big to be zipped.

Platform Specific

To build only on a specific platform use the property javacppPlatform:

# builds with support for all platforms
gradlew.bat releaseProcessingLib -PjavacppPlatform=linux-x86_64,macosx-x86_64,macosx-arm64,windows-x86_64,linux-armhf,linux-arm64

FAQ

Why is xy network not implemented?

Please open an issue if you have a cool network that could be implemented or just contribute a PR.

Why is it no possible to train my own network?

The idea was to give artist and makers a simple tool to run networks inside of Processing. To train a network needs a lot of specific knowledge about Neural Networks (CNN in specific).

Of course it is possible to train your own YOLO or SSDMobileNet and use the weights with this library. Check out the following example for detection facemasks: cansik/yolo-mask-detection

Is it compatible with Greg Borensteins OpenCV for Processing?

No, OpenCV for Processing uses the direct OpenCV Java bindings instead of JavaCV. Please only include either one library, because Processing gets confused if two OpenCV packages are imported.

About

Maintained by cansik with the help of the following dependencies:

Stock images from the following peoples have been used: