KITTI Distance

Overview

KITTI Distance asks how far away the closest vehicle (car, van or truck) is in a single front-camera frame from the KITTI driving benchmark. Labels come from KITTI’s 3D object annotations: the smallest camera-frame depth z over all vehicles is binned into four classes, <=8m, 8-20m, >20m and no_vehicle. This is the kitti/closest_vehicle_distance task of VTAB-1k and the “KITTI Distance” dataset in CLIP’s evaluation suite. Unlike most classification benchmarks it probes spatial/depth understanding rather than object identity.

The 7,481 labelled frames of KITTI’s object training set are split by drive (no drive appears in two splits), reproducing the TFDS kitti v3 split used by VTAB:

• Train: 6,347 images

• Validation: 423 images

• Test: 711 images

Data Structure

When accessing an example using ds[i], you will receive a dictionary with the following keys:

Key	Type	Description
image	PIL.Image.Image	~375×1242×3 RGB image (left colour camera)
label	int	0: <=8m, 1: 8-20m, 2: >20m, 3: no_vehicle
distance	float	Depth of the closest vehicle in metres (1000.0 if there is none)

Usage Example

from stable_datasets.images import KITTIDistance

# First run downloads ~12 GB of KITTI images (+ labels and devkit) and prepares the cache.
ds = KITTIDistance(split="train")
ds_all = KITTIDistance(split=None)  # train / validation / test

sample = ds[0]
print(sample.keys())  # {"image", "label", "distance"}

ds_torch = ds.with_format("torch")

References

• Official website: https://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d

• Task definition: https://github.com/google-research/task_adaptation (kitti/closest_vehicle_distance)

• License: CC BY-NC-SA 3.0

Citation

@inproceedings{geiger2012kitti,
    title={Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite},
    author={Geiger, Andreas and Lenz, Philip and Urtasun, Raquel},
    booktitle={Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2012}}

@article{zhai2019vtab,
    title={A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark},
    author={Zhai, Xiaohua and Puigcerver, Joan and Kolesnikov, Alexander and others},
    journal={arXiv preprint arXiv:1910.04867},
    year={2019}}