chocoluffy / firis

Let people act in the movies.

train_act.sh:
- TRAIN_LOGDIR="${WORK_DIR}/${DATASET_DIR}/${ACT_FOLDER}/${EXP_FOLDER}/train 19-2-11 更新 - TRAIN_LOGDIR="${WORK_DIR}/${DATASET_DIR}/${ACT_FOLDER}/${EXP_FOLDER}/train-19-2-12" 19-2-12 更新

ACT/parse-data.ipynb: parse supervisely dataset.
- draw segmentation polygon from opencv detection points.
- save masks.

ACT/convert_dataset.sh: convert dataset into TF-Records, similar to processing PASCAL VOC 2012 dataset.

Tips

• Run python2 on the scripts, to avoid type conversion bugs!
• tensorboard --logdir=${PATH_TO_LOG_DIRECTORY} --host=127.0.0.1 to visualize network architecture.
• Model backbone: DeepLab V3, reference on "Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs".