/tensorflow-object-contour-detection

A tensorflow implementation of object-contour-detection with fully convolutional encoder decoder network

Primary LanguagePythonMIT LicenseMIT

tensorflow-object-contour-detection

This is a tensorflow implimentation of Object Contour Detection with a Fully Convolutional Encoder-Decoder Network (https://arxiv.org/pdf/1603.04530.pdf) .

REQUIREMENTS :

pip install requirements.txt

Label Preparation :

To prepare the labels for contour detection from PASCAL Dataset , run create_lables.py and edit the file to add the path of the labels and new labels to be generated . Use this path for labels during training.

TRAINING :

python train.py \
    --max_to_keep=50 \
    --Epochs=100 \
    --momentum=0.9 \
    --learning_rate=.0000001 \
    --train_crop_size=480 \
    --clip_by_value=1.0 \
    --train_text = ${path to text file} \
    --log_dir = ${path to where logs will be saved} \
    --tf_initial_checkpoint=${PATH_TO_CHECKPOINT} \
    --label_dir = ${path to label directory} \
    --image_dir = ${path to image directory}

EVALUATION :

python eval.py \
    --checkpoint=${path to checkpoint to be evaluated} \
    --save_preds=${path to folder where predictions will be saved} \
    --image_dir = ${path to image directory} \
    --eval_crop_size=480 \
    --eval_text =  ${path to eval text file}

Results :

Image_1

prediction_1

Image_1

prediction_1