vyaslkv/keras-retinanet-logo

Keras implementation of RetinaNet object detection.

Keras RetinaNet Logo Detection

Keras implementation of RetinaNet object detection on logo detection. Forked on https://github.com/fizyr/keras-retinanet. Original paper is Focal Loss for Dense Object Detection.

Preparation

1. Clone this repository.
2. In the repository keras-retinanet execute python setup.py install --user. Please make sure tensorflow is installed as per your systems requirements. Also, make sure Keras 2.1.2 is installed.
3. This repository requires the master branch of keras-resnet (run pip install --user --upgrade git+https://github.com/broadinstitute/keras-resnet).

Training

1. Make sure to complete Preparation steps first.
2. Download FlickrLogos32 dataset from here. Extract FlickrLogos-v2 folder.
3. Download Logos-32plus_v1.0.1.zip dataset from here. Extract images folder and groundtruth.mat file to Logos32plus folder.
4. In main repository run command python prepare.py -f ./../FlickrLogos-v2/ -l ./../Flickr32plus/ -c ./csvpaths/classes.csv -t ./csvpaths/retina-train.csv -v ./csvpaths/retina-valid.csv -s ./csvpaths/retina-test.csv. Make sure -f option is FlickrLogos dataset folder path, -l is Logos32plus dataset path. In csvpaths folder files retina-valid.csv, retina-train.csv and retina-test.csv should have appeared.
5. Now run python train.py -n name_of_snapshot_folder -c ./csvpaths/classes.csv -t ./csvpaths/retina-train.csv -v ./csvpaths/retina-valid.csv. Folder with weighs and metadata should appear in snapshots folder.

Evaluating classification model

1. Make sure to complete Preparation steps first. Make sure to do 2-4 Training steps.
2. Train your own model or download weights from here.
3. In this repository run command python evaluate.py -w weights.h5 -c ./csvpaths/classes.csv -t ./csvpaths/retina-test.csv -o ./evalkit/classification.txt. -w is path to weights and -o is output path. In evalkit folder classification.txt file should appear.
4. In evalkit folder run command python fl_eval_classification.py --flickrlogos=..\..\FlickrLogos-v2 --classification="classification.txt". Make sure --flickrlogos option is path to FlickLogos32 dataset and --classification option is txt file from step 3. You can use original-classification.txt which is made on default weights.

Evaluating single photo or video

1. Make sure to complete Preparation steps first.
2. Train your own model or download weights from here.
3. To evaluate photo run command python test.py -f ./examples/test.png -o ./examples/output.png -w weights.h5 -c ./csvpaths/classes.csv. Where -f is your photo, -o is output photo, -w is weights.
4. To evaluate video run command python test_video.py -f ./examples/video.mp4 -o ./examples/output_video.mp4 -w weights.h5 -c ./csvpaths/classes.csv.

Dependencies

1. Tensorflow (https://www.tensorflow.org/install/).
2. Keras 2.1.2 pip install keras install after Tensorflow.
3. OpenCV pip install opencv-python.