Skip to content

Latest commit

 

History

History
115 lines (79 loc) · 3.23 KB

README.md

File metadata and controls

115 lines (79 loc) · 3.23 KB
Raw

Preprocess ScanNet data for path planning in indoor scenes

You can use our pre-processed data if you don't want to process it by yourself.

The data is organized as follows:

-| scannet_path_planning
---| height
-----| scene0000_00.pkl
-----| ...
---| mesh
-----| scene0000_00_vh_clean_2.ply # original ScanNet-V2 mesh
-----| ...
---| path
-----| scene0000_00.pkl
-----| ...
---| scene
-----| scene0000_00.npy
-----| ...

Data preparation

Download scannet dataset, and use the scripts in https://github.com/Silverster98/point_transformer.scannet to preprocess the scene.

Notes: you can only preprocess the selected scene, which is listed in preprocessing/scannet/scannet_annotate.py.

Annotate navigation graph mannually

By running:

python scannet_annotate.py

This script will provide a UI for user to annotate graph. Following the below tips to use this tool.

Tips:

  1. Click the canvas created by Matportlib, and this program will add a graph node in the scene.

  2. Continue to click, the nodes will connect in sequence, which means the progress will add edges.

  3. Press 'space' key, the progress will start a new path, the previous path will be stored. After creating several paths, all the paths will grouped as a graph.

    Note: if you click the mouse at a position which is very close to a previous node's position, then this click will use the previous node.

  4. Press 'd' to finish current scene

The annotated results will be saved in ./preprocessing/scannet/graph/ by default. Each *.pkl contains a dict data can be used to recover the graph. We provide the pre-annotated scene graph.

Generate height map for each scene

Execute:

python scannet_height.py --out_dir PATH_TO_SAVE

The height map are used for computing collision.

Each height/{scan_id}.pkl contains a floor height map used for computing collision. The format is

{
    'dim': int,             # height map dimension
    'height': np.ndarray,   # height map
    'minx': float,          # bounding box minx
    'maxx': float,          # bounding box maxx
    'miny': float,          # bounding box miny
    'maxy': float,          # bounding box maxy
}

Generate path for constructing training data

Execute:

python scannet_gen_path.py --out_dir PATH_TO_SAVE

This part will generate training data, i.e., path in scene, for path planning task. The generated data are formated as following:

- PATH_TO_SAVE/
    - path/
        - {scan_id}.pkl
        - ...

Each {scan_id}.pkl has the following format, storing a path list. Each element in this list is a tuple consisting a coarse path list and a numpy array. The coarse path list is a node id list within the original navigation graph. The numpy array with shape <N, 3> stores a refined path (position in space) that is processed from the coarse path. We can only use the refined path for training.

[
    ([node_id_1, node_id_2, ...], np.ndarray),
    ([node_id_1, node_id_2, ...], np.ndarray),
    ...
]

Train and test split

Train split: scene id less than 600

Test split: scene id greater than 600