DeepMAC model

DeepMAC (Deep Mask heads Above CenterNet) is a neural network architecture that is designed for the partially supervised instance segmentation task. For details see the The surprising impact of mask-head architecture on novel class segmentation paper. The figure below shows improved mask predictions for unseen classes as we use better mask-head architectures.

Just by using better mask-head architectures (no extra losses or modules) we achieve state-of-the-art performance in the partially supervised instance segmentation task.

Code structure

• deepmac_meta_arch.py implements our main architecture, DeepMAC, on top of the CenterNet detection architecture.
• The proto message DeepMACMaskEstimation in center_net.proto controls the configuration of the mask head used.
• The field allowed_masked_classes_ids controls which classes recieve mask supervision during training.
• Mask R-CNN based ablations in the paper are implemented in the TF model garden code base.

Prerequisites

1. Follow TF2 install instructions to install Object Detection API.
2. Generate COCO dataset by using create_coco_tf_record.py

Configurations

We provide pre-defined configs which can be run as a TF2 training pipeline. Each of these configurations needs to be passed as the pipeline_config_path argument to the object_detection/model_main_tf2.py binary. Note that the 512x512 resolution models require a TPU v3-32 and the 1024x1024 resolution models require a TPU v3-128 to train. The configs can be found in the configs/tf2 directory. In the table below X->Y indicates that we train with masks from X and evaluate with masks from Y. Performance is measured on the coco-val2017 set.

Partially supervised models

Fully supervised models

Here we report the Mask mAP averaged over all COCO classes on the test-dev2017 set .

Demos

• DeepMAC Colab lets you run a pre-trained DeepMAC model on user-specified boxes. Note that you are not restricted to COCO classes!
• iWildCam Notebook to visualize instance masks generated by DeepMAC on the iWildCam dataset.

Pre-trained models on COCO

Both these models take Image + boxes as input and produce per-box instance masks as output.

• CenterNet Hourglass backbone
• Mask-RCNN SpineNet backbone

See also

• Mask RCNN code in TF Model garden code base.
• Project website - git.io/deepmac

Citation

@misc{birodkar2021surprising,
      title={The surprising impact of mask-head architecture on novel class segmentation},
      author={Vighnesh Birodkar and Zhichao Lu and Siyang Li and Vivek Rathod and Jonathan Huang},
      year={2021},
      eprint={2104.00613},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}