Pytorch KITTI

This project aims to provide a simple yet effective way to scaffold and load the KITTI Vision Banchmark Dataset providing

• Datasets: Pytorch datasets to load each dataset

• Scaffolding: to download the datasets

• Metrics: common metrics used for each dataset

• Transformations: utilities to manipulate samples

Installation

To install torch-kitti

$ pip install torch-kitti

Scaffolding datasets

To manually download the datasets the torch-kitti command line utility comes in handy:

$ torch_kitti download --help
usage: Torch Kitti download [-h]
                            {sync_rectified,depth_completion,depth_prediction}
                            path

positional arguments:
  {sync_rectified,depth_completion,depth_prediction}
                        name of the dataset to download
  path                  where scaffold the dataset

optional arguments:
  -h, --help            show this help message and exit

Actually available datasets are:

• KITTI Depth Completion Dataset
• KITTI Depth Prediction Dataset
• KITTI Raw Sync+Rect Dataset

Loading Datasets

All datasets return dictionaries, utilities to manipulate them can be found in the torch_kitti.transforms module. Often each dataset provides options to include optional fields, for instance KittiDepthCompletionDataset usually provides simply the img, its sparse depth groundtruth gt and the sparse lidar hints lidar but using load_stereo=True stereo images will be included for each example.

from torchvision.transforms import Compose, RandomCrop, ToTensor

from torch_kitti.depth_completion import KittiDepthCompletionDataset
from torch_kitti.transforms import ApplyToFeatures

transform = ApplyToFeatures(
    Compose(
        [
            ToTensor(),
            RandomCrop([256, 512]),
        ]
    ),
    features=["img", "gt", "lidar"],
)

ds = KittiDepthCompletionDataset(
    "kitti_raw_sync_rect_root",
    "kitti_depth_completion_root",
    load_stereo=False,
    load_sequence=3,
    transform=transform,
    download=True,  # download if not found
)

Customizing Datasets

Each dataset exposes the elems attribute, containing the paths used by each example loaded, and it can be modified to customize the loaded data, it is composed by DataGroups each containing many DataElems. These latter can load many different file types and each loader automatically will load them if added to each DataGroup, the supported data types are image, depth, pcd, calib, imu, rt and intrinsics.

For example to load only a specific drive sequence for depth completion you could:

class SeqKittiDepthCompletionDataset(KittiDepthCompletionDataset):
    def __init__(self, drive_name: str, *args, **kwargs):
        super().__init__(*args, subset="all", **kwargs)
        self.elems = sorted(
            filter(lambda group: group.drive == drive_name and group.cam == 2, self.elems),
            key=lambda group: group.idx
        )

Contributing

Developing setup

Download from kitti and cd in the folder then prepare a virtual environment (1), install dev and doc dependencies (2) and pre-commit (3).

$ git clone https://github.com/andreaconti/torch_kitti.git
$ cd torch_kitti
$ python3 -m virtualenv .venv && source .venv/bin/activate  # (1)
$ pip install .[dev, doc] # (2)
$ pre-commit install  # (3)
$ python3 setup.py develop
$ pytest

Feel free to open an issue on GitHub, fork the repository and submit a pull request to solve bugs, improve docs, add datasets and features. All new feature must be tested.

Disclaimer on KITTI Vision Benchmark Suite

This library is an utility that downloads and prepares the dataset. The KITTI Vision Benchmark Suite is not hosted by this project nor it's claimed that you have license to use the dataset, it is your responsibility to determine whether you have permission to use this dataset under its license. You can find more details here.