Skip to content

Latest commit

 

History

History
113 lines (86 loc) · 5.61 KB

DATA.md

File metadata and controls

113 lines (86 loc) · 5.61 KB

⬇️ Data

Data Download

Data Preparation + Processing

We list the available data used in the current version of CrossOver in the table below:

Dataset Name Object Modality Scene Modality Object Temporal Information Scene Temporal Information
ScanNet [point, rgb, cad, referral] [point, rgb, floorplan, referral] ❌ ✅
3RScan [point, rgb, referral] [point, rgb, referral] ✅ ✅

We detail data download and release instructions for preprocessing with scripts for ScanNet + 3RScan.

  • For dataset download + data preparation, please look at README.MD in prepare_data/ directory.

Generated Embedding Data

We release the scene level embeddings created with CrossOver on the currenly used datasets in GDrive, which can be used for cross-modal retrieval with custom data as detailed in demo section.

  • embed_scannet.pt: Scene Embeddings For All Modalities (Point Cloud, RGB, Floorplan, Referral) in ScanNet
  • embed_scan3r.pt : Scene Embeddings For All Modalities (Point Cloud, RGB, Referral) in 3RScan

You agree to the terms of ScanNet, 3RScan, ShapeNet, Scan2CAD and SceneVerse datasets by downloading our hosted data.

File structure below:

{
  "scene": [{
    "scan_id": "the ID of the scan",
    "scene_embeds": {
        "modality_name"     : "modality_embedding"
      }
    "mask" : "modality_name" : "True/False whether modality was present in the scan"
    },
    {
      ...
    },...
  ]
}

🔧 Data Preprocessing

In order to process data faster during training + inference, we preprocess 1D (referral), 2D (RGB + floorplan) & 3D (Point Cloud + CAD) for both object instances and scenes. Note that, since for 3RScan dataset, they do not provide frame-wise RGB segmentations, we project the 3D data to 2D and store it in .pt format for every scan. We provide the scripts for projection. Here's an overview which data features are precomputed:

  • Object Instance: Referral, Multi-view RGB images, Point Cloud & CAD (only for ScanNet)
  • Scene: Referral, Multi-view RGB images, Floorplan (only for ScanNet) Point Cloud

We provide the preprocessing scripts which should be easily cusotmizable for new datasets. Further instructions below.

ScanNet

Running preprocessing scripts

Adjust the path parameters of Scannet in the config files under configs/preprocess (remember to adjust the path of Scannet:shape_dir to unzipped ShapeNet directory). Run the following (after changing the --config-path in the bash file):

$ bash scripts/preprocess/process_scannet.sh

Post running preprocessing, the data structure should look like the following:

Scannet/
├── objects_chunked/ (object data chunked into hdf5 format for instance baseline training)
|   ├── train_objects.h5
|   └── val_objects.h5
├── scans/
|   ├── scene0000_00/
|   │   ├── data1D.pt -> all 1D data + encoded (object referrals + BLIP features) 
|   │   ├── data2D.pt -> all 2D data + encoded (RGB + floorplan + DinoV2 features)
|   │   ├── data3D.pt -> all 3D data + encoded (Point Cloud + I2PMAE features - object only)
|   │   ├── object_id_to_label_id_map.pt -> Instance ID to NYU40 Label mapped
|   │   ├── objectsDataMultimodal.pt -> object data combined from data1D.pt + data2D.pt + data3D.pt (for easier loading)
|   │   ├── sel_cams_on_mesh.png (visualisation of the cameras selected for computing RGB features per scan)
|   │   ├── floor+obj.png -> rasterized floorplan (top-down projection of the floor+obj.ply)
|   |   └── floor+obj.ply -> floorplan + CAD mesh
|   └── ...

3RScan

Running preprocessing scripts

Adjust the path parameters of Scan3R in the config files under configs/preprocess. Run the following (after changing the --config-path in the bash file):

$ bash scripts/preprocess/process_scan3r.sh

Our script for 3RScan dataset performs the following additional processing:

  • 3D-to-2D projection for 2D segmentation and stores as gt-projection-seg.pt for each scan.

Post running preprocessing, the data structure should look like the following:

Scan3R/
├── objects_chunked/ (object data chunked into hdf5 format for instance baseline training)
|   ├── train_objects.h5
|   └── val_objects.h5
├── scans/
|   ├── 7f30f36c-42f9-27ed-87c6-23ceb65f1f9b/
|   │   ├── gt-projection-seg.pt -> 3D-to-2D projected data  consisting of framewise 2D instance segmentation
|   │   ├── data1D.pt -> all 1D data + encoded (object referrals + BLIP features) 
|   │   ├── data2D.pt -> all 2D data + encoded (RGB + floorplan + DinoV2 features)
|   │   ├── data2D_all_images.pt (RGB features of every image of every scan -- for comparison with SceneGraphLoc)
|   │   ├── data3D.pt -> all 3D data + encoded (Point Cloud + I2PMAE features - object only)
|   │   ├── object_id_to_label_id_map.pt -> Instance ID to NYU40 Label mapped
|   │   ├── objectsDataMultimodal.pt -> object data combined from data1D.pt + data2D.pt + data3D.pt (for easier loading)
|   │   └── sel_cams_on_mesh.png (visualisation of the cameras selected for computing RGB features per scan)
|   └── ...