Skip to content

Latest commit

 

History

History
192 lines (148 loc) · 7.39 KB

DATASETS.md

File metadata and controls

192 lines (148 loc) · 7.39 KB
Raw

Datasets

Please find the processed datasets used in our paper at here, including:

  1. Additional co-visibility masks and keypoint annotations for Nerfies-HyperNeRF dataset.
  2. Our accompanying iPhone dataset with more diverse and complex real-life motions.

All datasets are assumed to be organized as <DYCHECK_ROOT>/datasets/<DATASET>/<SEQUENCE>.

(0) Additional data

We follow and extend the data format adopted by Nerfies for better training and evaluation.

In total, we have the following (additional data are marked with **):

<DYCHECK_ROOT>/datasets/<DATASET>/<SEQUENCE>/
    ├── camera/
    │   └── <ITEM_ID>.json
    ├── **covisible/**
    │   └── <SCALE>x
    │       └── <SPLIT>
    │           └── <ITEM_ID>.png
    ├── **depth/**
    │   └── <SCALE>x
    │       └── <ITEM_ID>.npy
    ├── **keypoint/**
    │   └── <SCALE>x
    │       └── <SPLIT>
    │           ├── <ITEM_ID>.json
    │           └── skeleton.json
    ├── rgb/
    │   └── <SCALE>x
    │       └── <ITEM_ID>.png
    ├── **splits/**
    │   └── <SPLIT>.json
    ├── dataset.json
    ├── **emf**.json
    ├── **extra**.json
    ├── metadata.json
    ├── points.npy
    └── scene.json
Click to see per-item explainations.

covisible/

  • Contains pre-computed co-visibility masks for masked image metrics.
  • Only available if there are multi-camera validation frames.
  • For each pixel in the test image <ITEM_ID>.png given by the <SPLIT>, store the binary value of whether it is co-visible during training. 1 means co-visible and 0 the otherwise.
  • As specified in our paper, we deem a test pixel not co-visible if it is seen in less than max(5, len(train_imgs) // 10) training images.

depth/

  • Contains metric Lidar depth from iPhone sensor in millimeter.
  • Only available for the training frames in our iPhone dataset.

keypoint/

  • Contains our manual keypoint annotations for correspondence evaluation.
  • Only available for 10 uniformly sampled training frames.
  • Each keypoint frame <ITEM_ID>.json contains a (J, 3) list for J keypoints, where each row contain the (x, y, v) information for the absolute keypoint position (x, y) and its visibility in this frame v in binary.
  • skeleton.json stores a serialization of the skeleton definition.

splits/

  • Contains data split dictionary for toggling between the teleporting setting of existing methods and the actual monocular (non-teleporting) setting.
  • iPhone dataset contains simple train/val splits, while Nerfies-HyperNeRF contains train_/val_<SETTING> where <SETTING> is one of ["mono", "intl", "common"].
  • They are defined, use Nerfies-HyperNeRF as an example, and will be automatically generated the first time you run the dataloader (by running any of our evaluation/training scripts).
  • Each <SPLIT>.json contains a dictionary of the following format:
{
    // The name of each frame; equivalent to <ITEM_ID>.
    "frame_names": ["left1_000000", ...],
    // The camera ID for indexing the physical camera.
    "camera_ids": [0, ...],
    // The time ID for indexing the time step.
    "time_ids": [0, ...]
}

emf.json

  • Contins pre-computed effective multi-view factors of the captured training video in the following format:
{
    // The training video split name.
    "train_intl": {
        // The Full EMF.
        "Omega": 7.377937316894531,
        // The Angular EMF in deg/sec.
        "omega": 212.57649834023107
    },
    ...
}

extra.json

  • Contins extra information for evaluation and visualizations in the following format:
{
    // Scene bounding box; computed from the original sequence in the
    // normalized coordinate.
    "bbox": [
        [
            -0.24016861617565155,
            -0.35341497925615145,
            -0.6217879056930542
        ],
        [
            0.2697989344596863,
            0.09984857437828297,
            0.42420864422383164
        ]
    ],
    // Video downsampling factor for training and evaluation.
    "factor": 4,
    // Video frame rate.
    "fps": 15,
    // The scene look-at point in the normalized coordinate.
    "lookat": [
        0.0018036571564152837,
        -0.07246068120002747,
        0.05924934893846512
    ],
    // The scene up vector in the normalized coordinate.
    "up": [
        -0.07553146034479141,
        -0.9961089491844177,
        0.045409008860588074
    ]
}
  • They are defined, use Nerfies-HyperNeRF as an example, and will be automatically generated the first time you run the dataloader (by running any of our evaluation/training scripts).

(1) Nerfies-HyperNeRF dataset

Due to their similar capture protocol and data format, we combine the datasets released from Nerfies and HyperNeRF and denote it "Nerfies-HyperNeRF" dataset.

Please refer to Nerfies's and HyperNeRF's release pages for their download instructions. Together, there should be 7 sequences in total.

We prepared a simple download script for you to set up data from scratch:

# Download Nerfies and HyperNeRF datasets and inject our additional data.
bash scripts/download_nerfies_hypernerf_datasets.sh <DRIVE_NAME>
Click to see additional details.
  • The original training sequences on Nerfies-HyperNeRF are generated with severe camera motion due to camera teleportation.
  • We focus on the non-teleporting setting by only using the left camera for training in each capture.
  • We stripe the vrig- prefix from the HyperNeRF sequence names since we only focus on the captures with validation camera rig. This is also taken care of in our download script above.

(2) iPhone dataset

In our work, we propose a new dataset captured with iPhone captured in the wild, featuring diverse and complex motions comparing to Nerfies-HyperNeRF dataset.

This dataset contains 7 multi-camera captures and 7 single-camera captures. These captures are made without camera teleportation as we slowly moves the camera around a moving scene. We intentionally include a variety of subjects: general objects, quadruped (our pets), and humans. For visualizations on our captures, please refer to our project page.

We use a third party iOS app (Record3D) for depth sensing.

We prepared a simple download script for you to set up data from scratch:

# Download iPhone dataset.
bash scripts/download_iphone_dataset.sh <DRIVE_NAME>
Click to see additional details.
  • The whole dataset is around 28 GB in total.
  • To get started, you might want to just download one sequence first. Make sure you download each sequence into datasets/iphone/<SEQUENCE> directory.

(3) Process your own captures following our procedure

Please refer to RECORD3D_CAPTURE.md for our documentation.