xugk / orcnn

Learning to See the Invisible: End-to-End Trainable Amodal Instance Segmentation

ORCNN in Detectron2

Learning to See the Invisible: End-to-End Trainable Amodal Instance Segmentation
Waiyu Lam Instructor: Yong Jae Lee

Occlusion-aware RCNN propose an all-in-one, end to end trainable multi-task model for semantic segmentation that simultaneously predicts amodal masks, visible masks, and occlusion masks for each object instance in an image in a single forward pass. On the COCO amodal dataset, ORCNN outperforms the current baseline for amodal segmentation by a large margin.

The amodal mask is defined as the union of the visible mask and the invisible occlusion mask of the object.
Person:

Bench:

In this repository, we provide the code to train and evaluate ORCNN. We also provide tools to visualize occlusion mask annotation and results.

Installation

See INSTALL.md.

Quick Start

Inference with Pre-trained Models

See Getting Started Amodal

Training & Evaluation & Visualization

See Getting Started ORCNN

License

Detectron2 is released under the Apache 2.0 license.

Citing ORCNN

@article{DBLP:journals/corr/abs-1804-08864,
  author    = {Patrick Follmann and
               Rebecca K{\"{o}}nig and
               Philipp H{\"{a}}rtinger and
               Michael Klostermann},
  title     = {Learning to See the Invisible: End-to-End Trainable Amodal Instance
               Segmentation},
  journal   = {CoRR},
  volume    = {abs/1804.08864},
  year      = {2018},
  url       = {http://arxiv.org/abs/1804.08864},
  archivePrefix = {arXiv},
  eprint    = {1804.08864},
  timestamp = {Mon, 13 Aug 2018 16:46:01 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-1804-08864.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}