ClusterNet

This is an official implementation of "Online Unsupervised Video Object Segmentation via Contrastive Motion Clustering" (IEEE TCSVT).

Prerequisites

The training and testing experiments are conducted using PyTorch 1.8.1 with a single NVIDIA TITAN RTX GPU with 24GB Memory.

• python 3.8
• pytorch 1.8.1
• torchvision 0.9.1

conda create -n ClusterNet python=3.8
conda activate ClusterNet
conda install pytorch==1.8.1 torchvision==0.9.1 cudatoolkit=10.2 -c pytorch

Other minor Python modules can be installed by running

pip install opencv-python einops

Datasets

• DAVIS₁₆: We perform online clustering and evaluation on the validation set. However, please download DAVIS₁₇ (Unsupervised 480p) to fit the code.
• FBMS: This dataset contains videos of multiple moving objects, providing test cases for multiple object segmentation.
• SegTrackV2: Each sequence contains 1-6 moving objects.

Following the evaluation protocol in CIS, we combine multiple objects as a single foreground and use the region similarity $\mathcal{J}$ to measure the segmentation performance for the FBMS and SegTrackV2. Binary Mask: [FBMS][SegTrackV2]

• Path configuration: Dataset path settings is --data_dir in main.py.

parser.add_argument('--data_dir', default=None, type=str, help='dataset root dir')

• The datasets directory structure will be as follows:

|--DAVIS2017
|   |--Annotations_unsupervised
|   |   |--480p
|   |--ImageSets
|   |   |--2016
|   |--Flows_gap_1_${flow_method}
|       |--Full-Resolution
|--FBMS
|   |--Annotations_Binary
|   |--Flows_gap_1_${flow_method}
|--SegTrackv2
    |--Annotations_Binary
    |--Flows_gap_1_${flow_method}

Precompute optical flow

• The optical flow is estimated by using the PWCNet, RAFT and FlowFormer. In datasets directory, the variable flow_method is PWC, RAFT and FlowFormer, respectively.

• The flows are resized to the size of the original image (same as Motion Grouping), with each input frame having a size of $480\times854 for the *DAVIS₁₆* and \480\times640$for the *FBMS* and *SegTrackV2*. We convert the optical flow to 3-channel images with the standard visualization used for the optical flow and normalize it to$[-1, 1]$, and use only the previous frames for the optical flow estimation in the online setting.

Train & Inference

To train the ClusterNet model on a GPUs, you can use:

bash scripts/main.sh

In the main.sh file, first activate your Python environment and set gpu_id and data_dir. Then set the hyperparameters batch_size, n_clusters, and threshold to 16, 30, and 0.1, respectively.

Outputs

The model files and checkpoints will be saved in ./checkpoints/${exp_id}.

.pth files with _${sequence_name} store the network weights that initialize our autoencoder to train on DAVIS₁₆ through the loss of optical flow reconstruction.

The segmentation results will be saved in ./results/${exp_id}. The evaluation criterion is the mean region similarity $\mathcal{J}$.

Optical flow prediction	Method	Mean $\mathcal{J}\uparrow$
PWC-Net	MG ClusterNet	63.7 67.9(+4.2)
RAFT	MG ClusterNet	68.3 72.0(+3.7)
FlowFormer	MG ClusterNet	70.3 75.4(+5.1)

Citation

If you find our work useful in your research please consider citing our paper!

@ARTICLE{ClusterNet,
  author={Xi, Lin and Chen, Weihai and Wu, Xingming and Liu, Zhong and Li, Zhengguo},
  journal={IEEE Transactions on Circuits and Systems for Video Technology}, 
  title={Online Unsupervised Video Object Segmentation via Contrastive Motion Clustering}, 
  year={2023}
}  

Contact

If you have any questions, please feel free to contact Lin Xi (xilin1991@buaa.edu.cn).

Acknowledgement

This project would not have been possible without relying on some awesome repos: Motion Grouping, PWCNet, RAFT and FlowFormer. We thank the original authors for their excellent work.