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.
This code is borrowed from Keras Implementation of this model at https://github.com/fizyr/keras-retinanet and updated to run on Stanford Drone Data Set
- Clone this repository.
- Ensure numpy is installed using
pip install numpy --user - In the repository, execute
pip install . --user. Note that due to inconsistencies with howtensorflowshould be installed, this package does not define a dependency ontensorflowas it will try to install that (which at least on Arch Linux results in an incorrect installation). Please make suretensorflowis installed as per your systems requirements. - Alternatively, you can run the code directly from the cloned repository, however you need to run
python setup.py build_ext --inplaceto compile Cython code first. - Optionally, install
pycocotoolsif you want to train / test on the MS COCO dataset by runningpip install --user git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI.
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:
keras_retinanet/bin/train.py --weights snapshots/resnet50_coco_best_v2.1.0.h5 csv train_annotations.csv labels.csv --val-annotations val_annotations.csv
Here
- weights: Path to the weights for initializing training
- csv indicates retinanet is trained on a custom data set
- train_annotations.csv is path to training annotations
- labels.csv are the labels in the format class_name, class_id with 0 reserved for background class
- val_annotations is path to validation annotations
I uploaded the images used for training and validation to the google drive link below. Please download the same: https://drive.google.com/drive/u/1/folders/1QpE_iRDq1hUzYNBXSBSnmfe6SgTYE3J4
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
The class name to ID mapping file should contain one mapping per line. Each line should use the following format:
class_name,id
For the Stanford Drone Data Set, the training annotations are in train_annotations.csv, validation annotations are in val_annoations.csv and labels are in labels.csv
To calculate mean average precision on the validation set, please run
keras_retinanet/bin/evaluate.py csv val_annotations.csv labels.csv snapshots/resnet50_csv_08_inference.h5 --convert-model
Here we pass the val_annotations, labels and path to the trained weights
To run inference on the trained model, first step is to convert the trained model to a format that can be used by inference. The command for this is:
keras_retinanet/bin/convert_model.py snapshots/resnet50_csv_08.h5 snapshots/resnet50_csv_08_inference.h5
Here first path is the path to the trained model and the second would be the path to the converted inference model
I created two notebooks from the original code that can be used to run inference on images and on videos.
The notebooks are:
- ResNet50RetinaNet-Image.ipynb : For inference on a batch of images
- ResNet50RetinaNet-Video.ipynb : For inference on a video
I ran inference on one of the videos from Stanford Drone Data Set and shared here as death_circle_output.mp4.
The notebook ResNet50RetinaNet-Error-Analysis.ipynb can be used for error analysis and can help visualize true positive, class mismatch, false positive and false negative predictions
I have also uploaded my trained model at https://drive.google.com/drive/u/1/folders/1QpE_iRDq1hUzYNBXSBSnmfe6SgTYE3J4