karthikloganlogan/sf-mask-rcnn

Synthetic RGB-D Fusion (SF) Mask R-CNN for Unseen Object Instance Segmentation

Synthetic RGB-D Fusion (SF) Mask R-CNN

Synthetic RGB-D Fusion (SF) Mask R-CNN for unseen object instance segmentation

S. Back, J. Kim, R. Kang, S. Choi and K. Lee. Segmenting unseen industrial components in a heavy clutter using rgb-d fusion and synthetic data. 2020 IEEE International Conference on Image Processing (ICIP). IEEE, 2020. [Paper] [Video]

SF Mask R-CNN

---

Unseen object instance segmentation performance on WISDOM dataset

Method	Input	Use Synthetic Data	Backbone	mask AP	box AP	Reference
SD Mask R-CNN	Depth	Yes (WISDOM)	ResNet-35-FPN	51.6	-	Danielczuk et al.
Mask R-CNN	RGB	No	ResNet-35-FPN	38.4	-	Danielczuk et al.
Mask R-CNN	RGB	No	ResNet-50-FPN	40.1	36.7	Ito et al.
D-SOLO	RGB	No	ResNet-50-FPN	42.0	39.1	Ito et al.
PPIS	RGB	No	ResNet-50-FPN	52.3	48.1	Ito et al.
Mask R-CNN	RGB	Yes (Ours)	ResNet-50-FPN	59.0	61.4	Ours
Mask R-CNN	Depth	Yes (Ours)	ResNet-50-FPN	59.6	60.4	Ours
SF Mask R-CNN (early fusion)	RGB-Depth	Yes (Ours)	ResNet-50-FPN	55.5	57.2	Ours
SF Mask R-CNN (late fusion)	RGB-Depth	Yes (Ours)	ResNet-50-FPN	58.7	59.0	Ours
SF Mask R-CNN (confidence fusion)	RGB-Depth	Yes (Ours)	ResNet-50-FPN	60.5	61.0	Ours

SF Mask R-CNN is an upgraded version of RGB-D fusion Mask R-CNN with a confidence map estimator [1]. The main differences from [1] are

• SF Mask R-CNN generates a self-attention map from RGB and inpainted depth (validity mask and raw depth were used in [1])
• This self-attention map is used as a confidence map; Thus, RGB and depth feature maps fused with spatial self-attention in four different scales.
• It was fined-tuned on WISDOM-REAL-Train (100 images) and evaluated on public unseen object instance segmentation dataset, WISDOM (The only custom industrial dataset was used previously)

Updates

• SF Mask R-CNN has been released (2020/02/18)
• Paper extended from [1] with the detailed description, synthetic data, and robot demo will be publicly available soon... Stay tuned!

Getting Started

---

Environment Setup

1. Setup anaconda environment

$ conda create -n sfmaskrcnn python=3.7
$ conda activate sfmaskrcnn
$ pip install torch torchvision
$ pip install imgviz tqdm tensorboardX pandas opencv-python imutils pyfastnoisesimd scikit-image pycocotools
$ pip install pyrealsense2 # for demo
$ conda activate sfmaskrcnn

2. Download the provided SF Mask R-CNN weights pre-trained on our custom dataset.

• rgb only
• depth only
• early fusion
• late fusion
• confidence fusion

3. Download the WISDOM-Real dataset [Link]

4. Set the path to the dataset and pretrained weights (You can put this into your bash profile)

$ export WISDOM_PATH={/path/to/the/wisdom-real/high-res/dataset}
$ export WEIGHT_PATH={/path/to/the/pretrained/weights}

Train

To train an SF Mask R-CNN (confidence fusion, RGB-noisy depth as input) on a synthetic dataset.

$ python train.py --gpu 0 --cfg rgb_noisydepth_confidencefusion

To fine-tune the SF Mask R-CNN on WISDOM dataset

$ python train.py --gpu 0 --cfg rgb_noisydepth_confidencefusion_FT --resume

Evaluation

To evaluate an SF Mask R-CNN (confidence fusion, RGB-noisy depth as input) on a WISDOM dataset

$ python eval.py --gpu 0 --cfg rgb_noisydepth_confidencefusion \
    --eval_data wisdom \
    --dataset_path $WISDOM_PATH \
    --weight_path $WEIGHT_PATH/SFMaskRCNN_ConfidenceFusion.tar 

Visualization

To visualize the inference results of SF Mask R-CNN on a WISDOM dataset

$ python inference.py --gpu 0 --cfg rgb_noisydepth_confidencefusion \
    --eval_data wisdom --vis_depth \
    --dataset_path $WISDOM_PATH \
    --weight_path $WEIGHT_PATH/SFMaskRCNN_ConfidenceFusion.tar 

Our custom synthetic dataset

$ python inference.py --gpu 0 --cfg rgb_noisydepth_confidencefusion \
    --eval_data synthetic --vis_depth \
    --dataset_path examples \
    --weight_path $WEIGHT_PATH/SFMaskRCNN_ConfidenceFusion.tar 

Demo with RealSense

To run real-time demo with realsense-d435

# SF Mask R-CNN (confidence fusion)
$ python demo.py --cfg rgb_noisydepth_confidencefusion \
    --weight_path $WEIGHT_PATH/SFMaskRCNN_ConfidenceFusion.tar 

# SF Mask R-CNN (early fusion)
$ python demo.py --cfg rgb_noisydepth_earlyfusion \
    --weight_path $WEIGHT_PATH/SFMaskRCNN_EarlyFusion.tar 


# SF Mask R-CNN (late fusion)
$ python demo.py --cfg rgb_noisydepth_latefusion \
    --weight_path $WEIGHT_PATH/SFMaskRCNN_LateFusion.tar 

Authors

• Seunghyeok Back seungback
• Raeyoung Kang raeyo
• Taewon Kim ailabktw
• Joosoon Lee joosoon

Citation

If you use our work in a research project, please cite our work:

[1] @inproceedings{back2020segmenting,
  title={Segmenting unseen industrial components in a heavy clutter using rgb-d fusion and synthetic data},
  author={Back, Seunghyeok and Kim, Jongwon and Kang, Raeyoung and Choi, Seungjun and Lee, Kyoobin},
  booktitle={2020 IEEE International Conference on Image Processing (ICIP)},
  pages={828--832},
  year={2020},
  organization={IEEE}
}