We the creators of ImageAI are glad to announce 2 new AI projects to provide state-of-the-art Generative AI, LLM and Image Understanding on your personal computer and servers.
Install Jarvis on PC/Mac to setup limitless access to LLM powered AI Chats for your every day work, research and generative AI needs with 100% privacy and full offline capability.
Visit https://jarvis.genxr.co to get started.
TheiaEngine, the next-generation computer Vision AI API capable of all Generative and Understanding computer vision tasks in a single API call and available via REST API to all programming languages. Features include
- Detect 300+ objects ( 220 more objects than ImageAI)
- Provide answers to any content or context questions asked on an image
- very useful to get information on any object, action or information without needing to train a new custom model for every tasks
- Generate scene description and summary
- Convert 2D image to 3D pointcloud and triangular mesh
- Semantic Scene mapping of objects, walls, floors, etc
- Stateless Face recognition and emotion detection
- Image generation and augmentation from prompt
- etc.
Visit https://www.genxr.co/theia-engine to try the demo and join in the beta testing today.
ImageAI provides the most simple and powerful approach to training custom object detection models
using the YOLOv3 architeture, which
which you can load into the imageai.Detection.Custom.CustomObjectDetection
class. This allows
you to train your own YOLOv3 or TinyYOLOv3 model on any set of images that corresponds to any type of objects of interest.
The training process generates a JSON file that maps the objects names in your image dataset and the detection anchors, as well as creates lots of models. In choosing the best model for your custom object detection task, an evaluateModel()
function has been provided to compute the mAP of your saved models by allowing you to state your desired IoU and Non-maximum Suppression values. Then you can perform custom
object detection using the model and the JSON file generated.
- 🔳 Preparing your custom dataset
- 🔳 Training on your custom Dataset
- 🔳 Evaluating your saved detection models' mAP
To train a custom detection model, you need to prepare the images you want to use to train the model. You will prepare the images as follows:
- Decide the type of object(s) you want to detect and collect about 200 (minimum recommendation) or more picture of each of the object(s)
- Once you have collected the images, you need to annotate the object(s) in the images. ImageAI uses the YOLO for image annotation. You can generate this annotation for your images using the easy to use LabelImg image annotation tool, available for Windows, Linux and MacOS systems. Open the link below to install the annotation tool. See: https://github.com/tzutalin/labelImg
- When you are done annotating your images, annotation .txt files will be generated for each image in your dataset. The annotation .txt file describes each or all of the objects in the image. For example, if each image your image names are image(1).jpg, image(2).jpg, image(3).jpg till image(z).jpg; the corresponding annotation for each of the images will be image(1).txt, image(2).txt, image(3).txt till image(z).txt.
- Once you have the annotations for all your images, create a folder for your dataset (E.g headsets) and in this parent folder, create child folders train and validation
- In the train folder, create images and annotations sub-folders. Put about 70-80% of your dataset of each object's images in the images folder and put the corresponding annotations for these images in the annotations folder.
- In the validation folder, create images and annotations sub-folders. Put the rest of your dataset images in the images folder and put the corresponding annotations for these images in the annotations folder.
- Once you have done this, the structure of your image dataset folder should look like below:
>> train >> images >> img_1.jpg (shows Object_1) >> images >> img_2.jpg (shows Object_2) >> images >> img_3.jpg (shows Object_1, Object_3 and Object_n) >> annotations >> img_1.txt (describes Object_1) >> annotations >> img_2.txt (describes Object_2) >> annotations >> img_3.txt (describes Object_1, Object_3 and Object_n) >> validation >> images >> img_151.jpg (shows Object_1, Object_3 and Object_n) >> images >> img_152.jpg (shows Object_2) >> images >> img_153.jpg (shows Object_1) >> annotations >> img_151.txt (describes Object_1, Object_3 and Object_n) >> annotations >> img_152.txt (describes Object_2) >> annotations >> img_153.txt (describes Object_1)
- You can train your custom detection model completely from scratch or use transfer learning (recommended for better accuracy) from a pre-trained YOLOv3 model. Also, we have provided a sample annotated Hololens and Headsets (Hololens and Oculus) dataset for you to train with. Download the pre-trained YOLOv3 model and the sample datasets in the link below.
Download dataset hololens-yolo.zip
here and pre-trained model yolov3.pt
here
Before you start training your custom detection model, kindly take note of the following:
- The default batch_size is 4. If you are training with Google Colab, this will be fine. However, I will advice you use a more powerful GPU than the K80 offered by Colab as the higher your batch_size (8, 16), the better the accuracy of your detection model.
Then your training code goes as follows:
from imageai.Detection.Custom import DetectionModelTrainer
trainer = DetectionModelTrainer()
trainer.setModelTypeAsYOLOv3()
trainer.setDataDirectory(data_directory="hololens-yolo")
trainer.setTrainConfig(object_names_array=["hololens"], batch_size=4, num_experiments=200, train_from_pretrained_model="yolov3.pt")
# In the above,when training for detecting multiple objects,
#set object_names_array=["object1", "object2", "object3",..."objectz"]
trainer.trainModel()
Yes! Just 6 lines of code and you can train object detection models on your custom dataset. Now lets take a look at how the code above works.
from imageai.Detection.Custom import DetectionModelTrainer
trainer = DetectionModelTrainer()
trainer.setModelTypeAsYOLOv3()
trainer.setDataDirectory(data_directory="hololens-yolo")
In the first line, we import the ImageAI detection model training class, then we define the model trainer in the second line, we set the network type in the third line and set the path to the image dataset we want to train the network on.
trainer.setTrainConfig(object_names_array=["hololens"], batch_size=4, num_experiments=200, train_from_pretrained_model="yolov3.pt")
In the line above, we configured our detection model trainer. The parameters we stated in the function as as below:
- num_objects : this is an array containing the names of the objects in our dataset
- batch_size : this is to state the batch size for the training
- num_experiments : this is to state the number of times the network will train over all the training images, which is also called epochs
- train_from_pretrained_model(optional) : this is to train using transfer learning from a pre-trained YOLOv3 model
trainer.trainModel()
When you start the training, you should see something like this in the console:
Generating anchor boxes for training images...
thr=0.25: 1.0000 best possible recall, 6.93 anchors past thr
n=9, img_size=416, metric_all=0.463/0.856-mean/best, past_thr=0.549-mean:
====================
Pretrained YOLOv3 model loaded to initialize weights
====================
Epoch 1/100
----------
Train:
30it [00:14, 2.09it/s]
box loss-> 0.09820, object loss-> 0.27985, class loss-> 0.00000
Validation:
15it [01:45, 7.05s/it]
recall: 0.085714 precision: 0.000364 mAP@0.5: 0.000186, mAP@0.5-0.95: 0.000030
Epoch 2/100
----------
Train:
30it [00:07, 4.25it/s]
box loss-> 0.08691, object loss-> 0.07011, class loss-> 0.00000
Validation:
15it [01:37, 6.53s/it]
recall: 0.214286 precision: 0.000854 mAP@0.5: 0.000516, mAP@0.5-0.95: 0.000111
.
.
.
.
Let us explain the details shown above:
Generating anchor boxes for training images...
thr=0.25: 1.0000 best possible recall, 6.93 anchors past thr
n=9, img_size=416, metric_all=0.463/0.856-mean/best, past_thr=0.549-mean:
====================
Pretrained YOLOv3 model loaded to initialize weights
====================
The above details signifies the following:
-
ImageAI autogenerates the best match detection anchor boxes for your image dataset.
-
A the pretrained yolov3.pt was loaded to initalize the weights used to train the model.
Epoch 1/100
----------
Train:
30it [00:14, 2.09it/s]
box loss-> 0.09820, object loss-> 0.27985, class loss-> 0.00000
Validation:
15it [01:45, 7.05s/it]
recall: 0.085714 precision: 0.000364 mAP@0.5: 0.000186, mAP@0.5-0.95: 0.000030
Epoch 2/100
----------
Train:
30it [00:07, 4.25it/s]
box loss-> 0.08691, object loss-> 0.07011, class loss-> 0.00000
Validation:
15it [01:37, 6.53s/it]
recall: 0.214286 precision: 0.000854 mAP@0.5: 0.000516, mAP@0.5-0.95: 0.000111
- The above signifies the progress of the training.
- For each experiment (Epoch), a number of metrics are computed. The important once fo chosing an accuate models is detailed below
- The bounding box loss
box loss
is reported and expected to drop as the training progresses - The object localization loss
object loss
is reported and expected to drop as the training progresses - The class loss
class loss
is reported and expected to drop as the training progresses. If the class loss persists at 0.0000, it's because your dataset has a single class. - The
mAP50
andmAP0.5-0.95
metrics are expected to increase. This signifies the models accuracy increases. There might be flunctuations in these metrics sometimes.
- The bounding box loss
- For each increase in the
mAP50
after an experiment, a model is saved in the hololens-yolo/models folder. The higher the mAP50, the better the model.
Once you are done training, you can visit the link below for performing object detection with your custom detection model and detection_config.json file.
Detection/Custom/CUSTOMDETECTION.md
We have provided full documentation for all ImageAI classes and functions. Find links below:
- Documentation - English Version https://imageai.readthedocs.io