jjdblast / dvsnet

Modified implementation for DVSNet based on Tensorflow

DVSNet

The major contributors of this repository include Yu-Syuan Xu, Hsuan-Kung Yang, Tsu-Jui Fu, and Chun-Yi Lee.

Introduction

Dynamic Video Segmentation Network (DVSNet) framework is presented to strike a balance between quality and efficiency for semantic video segmentation. The DVSNet framework consists of two convolutional neural networks: a segmentation network (e.g., DeepLabv2) and a flow network (e.g.,FlowNet2). The former generates highly accurate semantic segmentations, but is deeper and slower. The latter is much faster than the former, but its output requires further processing to generate less accurate semantic segmentations. And DVSNet exploits a decision network (DN) to determine which frame regions should be forwarded to which paths based on a metric called expected confidence score. The usage of DN proposed an adaptive key frame scheduling policy to adaptively adjust the update period of key frames at runtime.

Demo Video

Disclaimer

This is a modified implementation for DVSNet based on Tensorflow. Please notes that there are some differences to the original implementation: (1) Data I/O and Image Preprocessing times are included when calculating fps. (2) It use NHWC data format rather than NCHW data format in the original implementation. These differences cause lower fps than reported in the paper.

Requirements

Checkpoint

Create checkpoint directory and get restore checkpoint from Google Drive.
Checkpoint all in one (DVSNet).
Checkpoint without decision network (finetune).

pip install in python 2.7

pip install tensorflow-gpu==1.4.1  # for Python 2.7 and GPU
pip install opencv-python
pip install Pillow
pip install scipy

Inference

To get segmented results of vedio frames:

python inference.py --data-dir=cityscape_video_dir --data-list=cityscape_video_list

List of Args:

--data_dir:      Path to the directory containing the dataset.
--data_list:     Path to the file listing the images in the dataset.
--restore_from:  Where restore model parameters from.
--decision_from: Where restore decision model parameters from (default same as restore_from).
--save_dir:      Where to save segmented output.
--num_steps:     Number of images in the video.
--overlap:       Overlapping size which must be dividable by 8.
--target:        Confidence score threshold.
--dynamic:       Whether to dynamically adjust target

Inference time including time of Data I/O and Image Preprocessing: 0.1~0.05s (10~20fps)
With Intel Xeon E5-2620 CPUs and NVIDIA GTX 1080 Ti GPU

Train

cd train/

step 1

Generate testcases (X=flow features, Y=confidence scores) for training decision network:

python gentestcase.py --data-dir=cityscape_dir --data-list=cityscape_list

List of Args:

--data_dir:     Path to the directory containing the dataset.
--data_list:    Path to the file listing the images in the dataset.
--restore_from: Where restore finetune(segmentation + flow) model parameters from.
--save_dir:     Where to save testcases.
--num_steps:    Number of generates testcases.
--clip:         Trim extreme testcases.

step 2

Train decision network:

python train.py --train-data-dir=train_testcase_dir --val-data-dir=val_testcase_dir

List of Args:

--train_data_dir: Path to the training testcases.
--val_data_dir:   Path to the validation testcases.
--save_dir:       Where to save decision model.
--batch_size:     Number of testcases sent to the network in one step.
--learning_rate:  Learning rate for training.
--epochs:         Number of epochs.
--decay:          Learning rate decay.

Citation

@inproceedings{xu2018dvsnet,
    author = {Yu-Shuan Xu and Hsuan-Kung Yang and Tsu-Jui Fu and Chun-Yi Lee},
    title = {Dynamic Video Segmentation Network},
    booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year = 2018
}

@article{chen2017deeplab,
    author = {L.-C. Chen and G. Papandreou and I. Kokkinos and K. Murphy and A. L. Yuille},
    title = {Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected {CRFs}},
    journal = {IEEE Trans. Pattern Analysis and Machine Intelligence (TPAMI)},
    year = 2017
}
@inproceedings{ilg2017flownet2,
    author = {E. Ilg and N. Mayer and T. Saikia and M. Keuper and A. Dosovitskiy and T. Brox},
    title = {FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks},
    booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year = 2017
}

Reference Code

DeepLabv2 tensorflow model : tensorflow-deeplab-resnet
FlowNet2 tensorflow model : flownet2-tf