keras implementation of Faster R-CNN, for image object detection of woman apparels in ecommerce products. The original code is from https://github.com/yhenon/keras-frcnn. Some pre-processing and model modification is required for your own dataset.
Training Datasets:
-
DeepFashion dataset "Fashion Landmark detection" dataset is used and can be downloaded from https://www.dropbox.com/sh/d4kj5owfoq1iio6/AACuupJTpw9Waw_Ri2uk60twa?dl=0. Extracting the image file requires author's authentication, please contact the author for password.
-
Artelab Dataset "Objects segmentation in the fashion field" dataset is used and can be downloaded from http://artelab.dista.uninsubria.it/downloads/datasets/fashion_field/object_segmentation/object_segmentation.html.
USAGE:
Both theano and tensorflow backends are supported. However compile times are very high in theano, and tensorflow is highly recommended.
requirements: keras version: 2.0.2 tensorflow: 1.2.1 numpy: 1.13.1 cv2: 3.2.0
Training:
We trained on the DeepFashion and Artelab Datasets separately, since the annotation file for the two datasets are in different format and requires different pre-processing codes.
- DeepFashion Dataset Before training, please download the dataset and the annotation file "list_bbox_consumer2shop.txt", and place in the root directory.
The full command is: python train_frcnn.py -o simple -p "list_bbox_consumer2shop.txt"
train_frcnn.py is used to train the DeepFashion model, with a pre-trained weights from ResNet50. -o shows the parser to preprocess the annotation file. -p indicates the image directory. The weights are saved in the format of "model_frcnn.hdf5_epoch_n" where n shows the epochs the weight is saved
preprocessing: the dataset contain images for both man's and woman's apparels. Since the testset contains only woman's apparels' images, we removed man's apparels' images (/img/CLOTHING).
- Artelab Dataset Before training, please download the dataset and the annotation file, and place in the directory of "artelab".
The full command is: python train.py -o artelab -p artelab/ObjectsSegmentationFashion_v1.0/
train.py is used to train the Artelab model, with a pre-trained weights from ResNet50. -o shows the parser to preprocess the annotation file. -p indicates the image directory. The weights are saved in the format of "artelab.hdf5_epoch_n" where n shows the epochs the weight is saved
Preprocessing: Man's apparels' images are removed, only 'drezzy_women_clothing' and 'drezzy_women_underwear' are used.
Prediction:
The two datasets cover different types of woman apparels, so the predictions are combined.
python test_combine.py -p "test"
-p indicates the directory where test images are placed.
The predicted images are labeled with bounding boxes and tags. They are saved in the "predicted_image" directory. 53 images from the dataset are selected for object detection.
Note: please replace the model_rpn.load_weights() and model_classifier.load_weights() with trained weights from the datasets.
Some well-detected exmaple outputs (Find more in the "predicted_images" file):
