dataset_prepare.md

Prepare datasets

It is recommended to symlink the dataset root to $mmdepth/data. If your folder structure is different, you may need to change the corresponding paths in config files.

mmdepth
├── data
│   ├── ddad 
│   ├── diml 
│   ├── DIODE 
│   ├── hypersim 
│   ├── ibims
│   ├── kitti 
│   ├── nyu 
│   ├── sunrgbd
│   └── vkitti2.0 

KITTI

Download the offical dataset from this link, including the raw data (about 200G) and fine-grained ground-truth depth maps.

Then, unzip the files into data/kitti. Remember to organizing the directory structure following instructions (Only need a few cut operations).

Finally, copy split files (whose names are started with kitti) in splits folder into data/kitti. Here, I utilize eigen splits following other supervised methods.

Some methods may use the camera intrinsic parameters (i.e., BTS), you need to download the benchmark_cam consisting of camera intrinsic parameters of the benchmark test set.

NYU-Depth V2

Following previous work, I utilize about 50K image-depth pairs as our training set and standard 652 images as the validation set. You can download the subset with the help of codes provided in BTS.

$ git clone https://github.com/cleinc/bts.git
$ cd bts
$ python utils/download_from_gdrive.py 1AysroWpfISmm-yRFGBgFTrLy6FjQwvwP sync.zip
$ unzip sync.zip

Then, you need to download the standard test set from this link. (Note: The downloaded file will be unzipped to folder test and train. You need to cut the files in the test folder out to data/nyu, organizing the directory structure following the file trees provided on the top of this page.)

Finally, copy nyu_train.txt and nyu_test.txt in the splits folder into the data/nyu.

SUNRGBD

The dataset can be download from this link.

Ibim-1 benchmark

The dataset can be download from this link.

DIODE

The dataset can be download from this link.

Hypersim

The dataset can be download from this link.

VKITTI2.0

The dataset can be download from this link.

DIML

The dataset can be download from this link.

Custom Dataset

We also provide a simple custom dataset class for users in depth/datasets/custom.py. Organize your data folder as our illustration. Note that instead of utilizing a split file to divide the train/val set, we directly classify data into train/val folder. A simple config file can be like:

train=dict(
    type=dataset_type,
    pipeline=dict(...),
    data_root='data/custom_dataset',
    test_mode=False,
    min_depth=1e-3,
    max_depth=10,
    depth_scale=1)

As for the custom dataset, we do not implement the evaluation details. If you want to get a quantitive metric result, you need to implement the pre_eval and evaluate functions following the ones in KITTI or other datasets.