Semantic-MapNet

Code for the paper:

Semantic MapNet: Building Allocentric Semantic Maps and Representations from Egocentric Views Vincent Cartillier, Zhile Ren, Neha Jain, Stefan Lee, Irfan Essa, Dhruv Batra

Website: smnet.com

Install

The code is tested with Ubuntu 16.04, Python 3.6, Pytorch v1.4+.

• Install the requirements using pip:

  pip install -r requirements.txt

• To render egocentric frames in the Matterport3D dataset we use the Habitat simulator. Install Habitat-sim and Habitat-lab: Tested with the following versions Habitat-sim == 0.1.7 and Habitat-lab == 0.1.6.

Demo

run the following script for demo:

python demo.py

Data

• data/paths.json has all the manually recorded trajectories.
• The semantic dense point cloud of objects with cleaned floor labels are available here: https://drive.google.com/drive/folders/1Fwbq7Bvl4kIjJ-YOJNbYWHD_6Gh8lFwQ?usp=sharing. Place those under data/object_point_clouds/. If you are looking to recompute those point clouds you can run data/build_point_cloud_from_mesh.py (will output a .ply file) or data/build_point_cloud_from_mesh_h5.py (will output a .h5 file / useful to compute the GT topdown map).
• Ground truth top-down semantic maps are available here: https://drive.google.com/drive/folders/1aM9vfDckY6K81mrVhVLmEX5rKZ2B1Q5r?usp=sharing. Place those under data/semmap/
• Place the Matterport3D data under data/mp3d/

Workflow

• To recompute the GT topdown semantic maps from the object point clouds (data/object_point_clouds/) you can run the following:

  python compute_GT_topdown_semantic_maps/build_semmap_from_obj_point_cloud.py

• Build training data: (1) build egocentric features + indices, (2) build topdown crops (250x250) (3) preprocess projection indices

  python precompute_training_inputs/build_data.py
  python precompute_training_inputs/build_crops.py
  python precompute_training_inputs/build_projindices.py

• To train SMNet you can run train.py

• Precompute testing features and projections indices for the full tours in the test set:

  python precompute_test_inputs/build_test_data.py
  python precompute_test_inputs/build_test_data_features.py

• To evaluate SMNet you can run test.py and:

  python eval/eval.py
  python eval/eval_bfscore.py

Pre-trained models

• pretrained weights are available here: https://drive.google.com/file/d/1KsJoTs91ez2bR35wW1VlD8jBG_gB-k7a/view?usp=sharing
• pretrained weights for RedNet are available here: https://drive.google.com/file/d/1PZDwl6dmIl6bhmWG42aRGyghgQQWTOcz/view?usp=sharing

Object-Goal Navigation

• Download the ObjectNav-Challenge-data and place it under: data/ObjectNav/objectnav_mp3d_v1/val/

• Download the precomputed topdown semantic map predictions here: https://drive.google.com/file/d/1wPtJaoDO15OtPcWcXuAbCtGQ3r-MdQM2/view?usp=sharing and place them in data/ObjectNav/semmap/

• Download the precomputed ObjNav GT goals here: https://drive.google.com/file/d/1Y6Qb6eGryZNkbjWGiqQJE2k-zV0aArrd/view?usp=sharing and place the json file in data/ObjectNav/

• You can recompute the semantic predictions using the explorations paths in data/ObjectNav/paths.json and the test.py script.

• Compute the freespace maps:

  python ObjectNav/build_freespace_maps.py

• Run A* path planning:

  python ObjectNav/run_astar_planning.py

Replica

• You can find the manually generated path related to the replica experiment at data_replica/paths.json

Citation

If you find our work useful in your research, please consider citing:

@article{cartillier2020semantic,
  title={Semantic MapNet: Building Allocentric SemanticMaps and Representations from Egocentric Views},
  author={Cartillier, Vincent and Ren, Zhile and Jain, Neha and Lee, Stefan and Essa, Irfan and Batra, Dhruv},
  journal={arXiv preprint arXiv:2010.01191},
  year={2020}
}

License

BSD