SpatialFlow: Bridging all Task for Panoptic Segmentation

SpatialFlow: Bridging all Task for Panoptic Segmentation,

Qiang Chen, Anda Cheng, Xiangyu He, Peisong Wang, Jian Cheng

In IEEE Transactions on Circuits and Systems for Video Technology,

arXiv preprint (arXiv 1910.08787)

Installation

• We adopt the codebase mmdetection (v2.3.0) with hash code 9596b9a.
• Please follow the guide to install mmdet.
• To prepare the dataset, please follow the guidance of dataset converters.

Modification for MMCV (v1.0.5 or v1.1.0)

• For v1.1.0, Comment out the line kwargs.setdefault('default', set_default) in the function dump_to_fileobj in mmcv to use MyJsonEncoder instead for mmcv.dump.
• Add a cregistry for ModulatedDeformConv2d in this line as follow

  @CONV_LAYERS.register_module('ModulatedDeformConv')
  class ModulatedDeformConv2d(nn.Module):
  ...

• Add support for ModulatedDeformConv2d in ConvModule here

  def forward(self, x, offset=None, mask=None, activate=True, norm=True):
  for layer in self.order:
      if layer == 'conv':
          if self.with_explicit_padding:
              x = self.padding_layer(x)
          if offset is not None:
              if mask is not None:
                  x = self.conv(x, offset, mask)
              else:
                  x = self.conv(x, offset)
          else:
              x = self.conv(x)
      elif layer == 'norm' and norm and self.with_norm:
          x = self.norm(x)
      elif layer == 'act' and activate and self.with_activation:
          x = self.activate(x)
  return x

Main results

Panoptic Segmentation on COCO validation

SpatialFlow-res50 (download)

SpatialFlow-res101 (download)

SpatialFlow-res101-DCN-MultiScale (download)

Panoptic Segmentation on COCO test-dev

SpatialFlow-res101-DCN-MultiScale (download)(test_dev_results)

Disclaimer:

• This is a reimplementation using mmdetv2 and the PQ is slightly higher than the one of our original paper (~0.9% PQ).
• We use fp16 training for res101-DCN-Multiscale.
• All the results listed here are measured on COCO val split. According to our experimental results, the performance on COCO test-dev will be slightly higher than the one on COCO val (~0.5% PQ).

Usage

• The learning_rate in configs is setting default for 4 GPUs.

Train with multiple GPUs

./tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM}

# Example: 
# ./tools/dist_train.sh configs/spatialflow/spatialflow_r50_fpn_20e_coco.py 4

Testing

# multi-gpu testing
./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} --out ${OUTPUT_FILE} --eval panoptic

Example: 
./tools/dist_test.sh configs/spatialflow/spatialflow_r50_fpn_20e_coco.py 
./work_dir/spatialflow_r50_fpn_20e_coco/latest.pth  8  --out ./work_dir/spatialflow_r50_fpn_20e_coco/results.pkl --eval panoptic

# single-gpu testing
python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} --out ${OUTPUT_FILE} --eval panoptic

Example: 
python tools/test.py configs/spatialflow/spatialflow_r50_fpn_20e_coco.py  ./work_dir/spatialflow_r50_fpn_20e_coco/latest.pth 
--out ./work_dir/spatialflow_r50_fpn_20e_coco/results.pkl --eval panootic

Citation

Please consider citing our papers in your publications if the project helps your research. BibTeX reference is as follows.

@artical{chen2020spatialflow,
  author={Chen, Qiang and Cheng, Anda and He, Xiangyu and Wang, Peisong and 
  Cheng, Jian},
  journal={IEEE Transactions on Circuits and Systems for Video Technology}, 
  title={SpatialFlow: Bridging all Task for Panoptic Segmentation}, 
  year={2020}
  doi={10.1109/TCSVT.2020.3020257}}