Keras RetinaNet

Keras implementation of RetinaNet object detection as described in Focal Loss for Dense Object Detection by Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He and Piotr Dollár.

Warning

This repository is migrating to (tf-retinanet)[https://github.com/fizyr/tf-retinanet], a tf.keras porting of keras-retinanet.

Installation

1. Clone this repository.
2. Ensure numpy is installed using pip install numpy --user
3. In the repository, execute pip install . --user. Note that due to inconsistencies with how tensorflow should be installed, this package does not define a dependency on tensorflow as it will try to install that (which at least on Arch Linux results in an incorrect installation). Please make sure tensorflow is installed as per your systems requirements.
4. Alternatively, you can run the code directly from the cloned repository, however you need to run python setup.py build_ext --inplace to compile Cython code first.
5. Optionally, install pycocotools if you want to train / test on the MS COCO dataset by running pip install --user git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI.

Testing

An example of testing the network can be seen in this Notebook. In general, inference of the network works as follows:

boxes, scores, labels = model.predict_on_batch(inputs)

Where boxes are shaped (None, None, 4) (for (x1, y1, x2, y2)), scores is shaped (None, None) (classification score) and labels is shaped (None, None) (label corresponding to the score). In all three outputs, the first dimension represents the shape and the second dimension indexes the list of detections.

Loading models can be done in the following manner:

from keras_retinanet.models import load_model
model = load_model('/path/to/model.h5', backbone_name='resnet50')

Execution time on NVIDIA Pascal Titan X is roughly 75msec for an image of shape 1000x800x3.

Converting a training model to inference model

The training procedure of keras-retinanet works with training models. These are stripped down versions compared to the inference model and only contains the layers necessary for training (regression and classification values). If you wish to do inference on a model (perform object detection on an image), you need to convert the trained model to an inference model. This is done as follows:

# Running directly from the repository:
keras_retinanet/bin/convert_model.py /path/to/training/model.h5 /path/to/save/inference/model.h5

# Using the installed script:
retinanet-convert-model /path/to/training/model.h5 /path/to/save/inference/model.h5

Most scripts (like retinanet-evaluate) also support converting on the fly, using the --convert-model argument.

Training

keras-retinanet can be trained using this script. Note that the train script uses relative imports since it is inside the keras_retinanet package. If you want to adjust the script for your own use outside of this repository, you will need to switch it to use absolute imports.

If you installed keras-retinanet correctly, the train script will be installed as retinanet-train. However, if you make local modifications to the keras-retinanet repository, you should run the script directly from the repository. That will ensure that your local changes will be used by the train script.

The default backbone is resnet50. You can change this using the --backbone=xxx argument in the running script. xxx can be one of the backbones in resnet models (resnet50, resnet101, resnet152), mobilenet models (mobilenet128_1.0, mobilenet128_0.75, mobilenet160_1.0, etc), densenet models or vgg models. The different options are defined by each model in their corresponding python scripts (resnet.py, mobilenet.py, etc).

Trained models can't be used directly for inference. To convert a trained model to an inference model, check here.

Usage

For training on Pascal VOC, run:

# Running directly from the repository:
keras_retinanet/bin/train.py pascal /path/to/VOCdevkit/VOC2007

# Using the installed script:
retinanet-train pascal /path/to/VOCdevkit/VOC2007

For training on MS COCO, run:

# Running directly from the repository:
keras_retinanet/bin/train.py coco /path/to/MS/COCO

# Using the installed script:
retinanet-train coco /path/to/MS/COCO

For training on Open Images Dataset OID or taking place to the OID challenges, run:

# Running directly from the repository:
keras_retinanet/bin/train.py oid /path/to/OID

# Using the installed script:
retinanet-train oid /path/to/OID

# You can also specify a list of labels if you want to train on a subset
# by adding the argument 'labels_filter':
keras_retinanet/bin/train.py oid /path/to/OID --labels-filter=Helmet,Tree

# You can also specify a parent label if you want to train on a branch
# from the semantic hierarchical tree (i.e a parent and all children)
(https://storage.googleapis.com/openimages/challenge_2018/bbox_labels_500_hierarchy_visualizer/circle.html)
# by adding the argument 'parent-label':
keras_retinanet/bin/train.py oid /path/to/OID --parent-label=Boat

For training on KITTI, run:

# Running directly from the repository:
keras_retinanet/bin/train.py kitti /path/to/KITTI

# Using the installed script:
retinanet-train kitti /path/to/KITTI

If you want to prepare the dataset you can use the following script:
https://github.com/NVIDIA/DIGITS/blob/master/examples/object-detection/prepare_kitti_data.py

For training on a [custom dataset], a CSV file can be used as a way to pass the data. See below for more details on the format of these CSV files. To train using your CSV, run:

# Running directly from the repository:
keras_retinanet/bin/train.py csv /path/to/csv/file/containing/annotations /path/to/csv/file/containing/classes

# Using the installed script:
retinanet-train csv /path/to/csv/file/containing/annotations /path/to/csv/file/containing/classes

In general, the steps to train on your own datasets are:

1. Create a model by calling for instance keras_retinanet.models.backbone('resnet50').retinanet(num_classes=80) and compile it. Empirically, the following compile arguments have been found to work well:

model.compile(
    loss={
        'regression'    : keras_retinanet.losses.smooth_l1(),
        'classification': keras_retinanet.losses.focal()
    },
    optimizer=keras.optimizers.adam(lr=1e-5, clipnorm=0.001)
)

2. Create generators for training and testing data (an example is show in keras_retinanet.preprocessing.pascal_voc.PascalVocGenerator).
3. Use model.fit_generator to start training.

Pretrained models

All models can be downloaded from the releases page.

MS COCO

Results using the cocoapi are shown below (note: according to the paper, this configuration should achieve a mAP of 0.357).

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.350
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.537
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.374
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.191
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.383
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.472
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.306
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.491
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.533
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.345
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.577
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.681

Open Images Dataset

There are 3 RetinaNet models based on ResNet50, ResNet101 and ResNet152 trained on all 500 classes of the Open Images Dataset (thanks to @ZFTurbo).

Backbone	Image Size (px)	Small validation mAP	LB (Public)
ResNet50	768 - 1024	0.4594	0.4223
ResNet101	768 - 1024	0.4986	0.4520
ResNet152	600 - 800	0.4991	0.4651

For more information, check @ZFTurbo's repository.

CSV datasets

The CSVGenerator provides an easy way to define your own datasets. It uses two CSV files: one file containing annotations and one file containing a class name to ID mapping.

Annotations format

The CSV file with annotations should contain one annotation per line. Images with multiple bounding boxes should use one row per bounding box. Note that indexing for pixel values starts at 0. The expected format of each line is:

path/to/image.jpg,x1,y1,x2,y2,class_name

By default the CSV generator will look for images relative to the directory of the annotations file.

Some images may not contain any labeled objects. To add these images to the dataset as negative examples, add an annotation where x1, y1, x2, y2 and class_name are all empty:

path/to/image.jpg,,,,,

A full example:

/data/imgs/img_001.jpg,837,346,981,456,cow
/data/imgs/img_002.jpg,215,312,279,391,cat
/data/imgs/img_002.jpg,22,5,89,84,bird
/data/imgs/img_003.jpg,,,,,

This defines a dataset with 3 images. img_001.jpg contains a cow. img_002.jpg contains a cat and a bird. img_003.jpg contains no interesting objects/animals.

Class mapping format

The class name to ID mapping file should contain one mapping per line. Each line should use the following format:

class_name,id

Indexing for classes starts at 0. Do not include a background class as it is implicit.

For example:

cow,0
cat,1
bird,2

Anchor optimization

In some cases, the default anchor configuration is not suitable for detecting objects in your dataset, for example, if your objects are smaller than the 32x32px (size of the smallest anchors). In this case, it might be suitable to modify the anchor configuration, this can be done automatically by following the steps in the anchor-optimization repository. To use the generated configuration check here for an example config file and then pass it to train.py using the --config parameter.

Debugging

Creating your own dataset does not always work out of the box. There is a debug.py tool to help find the most common mistakes.

Particularly helpful is the --annotations flag which displays your annotations on the images from your dataset. Annotations are colored in green when there are anchors available and colored in red when there are no anchors available. If an annotation doesn't have anchors available, it means it won't contribute to training. It is normal for a small amount of annotations to show up in red, but if most or all annotations are red there is cause for concern. The most common issues are that the annotations are too small or too oddly shaped (stretched out).

Results

MS COCO

Status

Example output images using keras-retinanet are shown below.

Projects using keras-retinanet

• Improving Apple Detection and Counting Using RetinaNet. This work aims to investigate the apple detection problem through the deployment of the Keras RetinaNet.
• Improving RetinaNet for CT Lesion Detection with Dense Masks from Weak RECIST Labels. Research project for detecting lesions in CT using keras-retinanet.
• NudeNet. Project that focuses on detecting and censoring of nudity.
• Individual tree-crown detection in RGB imagery using self-supervised deep learning neural networks. Research project focused on improving the performance of remotely sensed tree surveys.
• ESRI Object Detection Challenge 2019. Winning implementation of the ESRI Object Detection Challenge 2019.
• Lunar Rockfall Detector Project. The aim of this project is to map lunar rockfalls on a global scale using the available > 1.6 million satellite images.
• NATO Innovation Challenge. The winning team of the NATO Innovation Challenge used keras-retinanet to detect cars in aerial images (COWC dataset).
• Microsoft Research for Horovod on Azure. A research project by Microsoft, using keras-retinanet to distribute training over multiple GPUs using Horovod on Azure.
• Anno-Mage. A tool that helps you annotate images, using input from the keras-retinanet COCO model as suggestions.
• Telenav.AI. For the detection of traffic signs using keras-retinanet.
• Towards Deep Placental Histology Phenotyping. This research project uses keras-retinanet for analysing the placenta at a cellular level.
• 4k video example. This demo shows the use of keras-retinanet on a 4k input video.
• boring-detector. I suppose not all projects need to solve life's biggest questions. This project detects the "The Boring Company" hats in videos.
• comet.ml. Using keras-retinanet in combination with comet.ml to interactively inspect and compare experiments.
• Weights and Biases. Trained keras-retinanet on coco dataset from beginning on resnet50 and resnet101 backends.
• Google Open Images Challenge 2018 15th place solution. Pretrained weights for keras-retinanet based on ResNet50, ResNet101 and ResNet152 trained on open images dataset.
• poke.AI. An experimental AI that attempts to master the 3rd Generation Pokemon games. Using keras-retinanet for in-game mapping and localization.
• retinanetjs A wrapper to run RetinaNet inference in the browser / Node.js. You can also take a look at the example app.

If you have a project based on keras-retinanet and would like to have it published here, shoot me a message on Slack.

Notes

• This repository requires Keras 2.3.0 or higher.
• This repository is tested using OpenCV 3.4.
• This repository is tested using Python 2.7 and 3.6.

Contributions to this project are welcome.

Discussions

Feel free to join the #keras-retinanet Keras Slack channel for discussions and questions.

FAQ

• I get the warning UserWarning: No training configuration found in save file: the model was not compiled. Compile it manually., should I be worried? This warning can safely be ignored during inference.
• I get the error ValueError: not enough values to unpack (expected 3, got 2) during inference, what to do?. This is because you are using a train model to do inference. See https://github.com/fizyr/keras-retinanet#converting-a-training-model-to-inference-model for more information.
• How do I do transfer learning? The easiest solution is to use the --weights argument when training. Keras will load models, even if the number of classes don't match (it will simply skip loading of weights when there is a mismatch). Run for example retinanet-train --weights snapshots/some_coco_model.h5 pascal /path/to/pascal to transfer weights from a COCO model to a PascalVOC training session. If your dataset is small, you can also use the --freeze-backbone argument to freeze the backbone layers.
• How do I change the number / shape of the anchors? The train tool allows to pass a configuration file, where the anchor parameters can be adjusted. Check here for an example config file.
• I get a loss of 0, what is going on? This mostly happens when none of the anchors "fit" on your objects, because they are most likely too small or elongated. You can verify this using the debug tool.
• I have an older model, can I use it after an update of keras-retinanet? This depends on what has changed. If it is a change that doesn't affect the weights then you can "update" models by creating a new retinanet model, loading your old weights using model.load_weights(weights_path, by_name=True) and saving this model. If the change has been too significant, you should retrain your model (you can try to load in the weights from your old model when starting training, this might be a better starting position than ImageNet).
• I get the error ModuleNotFoundError: No module named 'keras_retinanet.utils.compute_overlap', how do I fix this? Most likely you are running the code from the cloned repository. This is fine, but you need to compile some extensions for this to work (python setup.py build_ext --inplace).
• How do I train on my own dataset? The steps to train on your dataset are roughly as follows:
• 1. Prepare your dataset in the CSV format (a training and validation split is advised).
• 2. Check that your dataset is correct using retinanet-debug.
• 3. Train retinanet, preferably using the pretrained COCO weights (this gives a far better starting point, making training much quicker and accurate). You can optionally perform evaluation of your validation set during training to keep track of how well it performs (advised).
• 4. Convert your training model to an inference model.
• 5. Evaluate your inference model on your test or validation set.
• 6. Profit!