HGPU

Code for paper: Hierarchical Graph Pattern Understanding for Zero-Shot Video Object Segmentation

Figure 1. Architecture of HGPU. Frames (I_k, I_{k+1}) and their corresponding optical flow (O_k) are used as inputs to generate high-order feature representations with multiple modalities by (a) HGPE. Motion-appearance features are parsed through (b) MAUD to output coarse-to-fine segmentation results.

Requirements

• Python (3.6.12)

• PyTorch (version:1.7.0)

• Apex (https://github.com/NVIDIA/apex, version: 0.1)

• Requirements in the requirements.txt files.

Training

Download Datasets

1. Download the DAVIS-2017 dataset from DAVIS.
2. Download the YouTube-VOS dataset from YouTube-VOS, the validation set is not mandatory to download.
3. Download the YouTube2018-hed and davis-hed datasets.
4. The optical flow files are obtained by RAFT, we provide demo code that can be run directly on path flow. We also provide optical flow of YouTube-VOS (18G) in GoogleDrive, optical flow of DAVIS can be found in GoogleDrive.

Dataset Format

Please ensure the datasets are organized as following format.

YouTube-VOS
|----train
      |----Annotations
      |----JPEGImages
      |----YouTube-flow
      |----YouTube-hed
      |----meta.json

DAVIS
      |----Annotations
      |----ImageSets
      |----JPEGImages
      |----davis-flow
      |----davis-hed

Run train.py

We provide multi-GPU parallel code based on apex. Run CUDA_VISIBLE_DEVICES="0,1,2,3" python -m torch.distributed.launch --nproc_per_node 4 train_HGPU.py for distributed training in Pytorch.

Note:

Please change your dada path in two codes (libs/utils/config_davis.py in line 52, and libs/utils/config_youtubevos.py in line 38)

Testing & Results

If you want to test the model results directly, you can follow the settings below.

1. Pre-trained models and pre-calculated results.

Table 1. A list of available pre-trained models and pre-calculated results.

Dataset	J Mean	Pre-trained Model	Pre-calculated Mask
DAVIS-16	86.0 full-training; 78.6 pre-training; 80.3 main-training	[full-training]; [pre-training]; [main-training]	[DAVIS-16]
Youtube-Objects	73.9 (average 10 category levels)	without fine-tuning	[YouTube-Objects]
Long-Videos	74.0	without fine-tuning	[Long-Videos]
DAVIS-17	67.0	[DAVIS-17]	[DAVIS-17]

2. Change your path in test_HGPU.py, then run python test_HGPU.py.

3. Run python apply_densecrf_davis.py (change your path in line 16) for getting binary masks.

The code directory structure is as follows.

  HGPU
  |----flow
  |----libs
  |----model
     |----HGPU
        |----encoder_0.8394088008745166.pt
        |----decoder_0.8394088008745166.pt
        |....
  |----apply_densecrf_davis.py
  |----train_HGPU.py
  |----test_HGPU.py

4. Evaluation code from DAVIS_Evaluation, the python version is available at PyDavis16EvalToolbox.

5. The YouTube-Objects dataset can be downloaded from here and annotation can be found here.

6. The Long-Videos dataset can be downloaded from here.

Supplementary

1. PWCNet is used to compute optical flow estimation for HGPU training, and the pre-trained model and inference results are in here.

Acknowledge

1. Zero-shot Video Object Segmentation via Attentive Graph Neural Networks, ICCV 2019 (https://github.com/carrierlxk/AGNN)
2. Motion-Attentive Transition for Zero-Shot Video Object Segmentation, AAAI 2020 (https://github.com/tfzhou/MATNet)
3. Video Object Segmentation Using Space-Time Memory Networks, ICCV 2019 (https://github.com/seoungwugoh/STM)
4. Video Object Segmentation with Adaptive Feature Bank and Uncertain-Region Refinement, NeurIPS 2020 (https://github.com/xmlyqing00/AFB-URR)