KITTI Object Detection Format¶

Overview¶

The KITTI format was developed as part of the KITTI Vision Benchmark Suite, focusing on autonomous driving scenarios. This format is particularly well-suited for 3D object detection and tracking tasks. The complete format specification can be found in the KITTI development kit documentation.

Specification of KITTI Detection Format¶

Each object is represented by 15 space-separated values:

#Values    Name      Description
----------------------------------------------------------------------------
   1    type         Object type (Car, Van, Truck, etc.)
   1    truncated    Float 0-1 (truncated ratio)
   1    occluded     Integer (0=visible, 1=partly occluded, 2=fully occluded)
   1    alpha        Observation angle (-pi..pi)
   4    bbox         2D bounding box (x1,y1,x2,y2) in pixels
   3    dimensions   3D dimensions (height, width, length) in meters
   3    location     3D location (x,y,z) in camera coordinates
   1    rotation_y   Rotation around Y-axis in camera coordinates

Directory Structure of KITTI Dataset¶

dataset/
├── images/
│   ├── 000000.png
│   └── 000001.png
└── labels/
    ├── 000000.txt
    └── 000001.txt

Label Format¶

# Example: 000000.txt
Car -1 -1 -10 614 181 727 284 -1 -1 -1 -1000 -1000 -1000 -10
Pedestrian -1 -1 -10 123 456 789 012 -1 -1 -1 -1000 -1000 -1000 -10

Note: The filename of each label file must match its corresponding image file, with .txt extension.

Annotation Format Conversion¶

Using CLI¶

Convert from YOLOv8 to KITTI format:

labelformat convert \
    --task object-detection \
    --input-format yolov8 \
    --input-file yolo-labels/data.yaml \
    --input-split train \
    --output-format kitti \
    --output-folder kitti-labels

Convert from KITTI to YOLOv8 format:

labelformat convert \
    --task object-detection \
    --input-format kitti \
    --input-folder kitti-labels \
    --category-names car,pedestrian,cyclist \
    --images-rel-path ../images \
    --output-format yolov8 \
    --output-file yolo-labels/data.yaml \
    --output-split train

Using Python¶

from pathlib import Path
from labelformat.formats import KittiObjectDetectionInput, YOLOv8ObjectDetectionOutput

# Load KITTI labels
label_input = KittiObjectDetectionInput(
    input_folder=Path("kitti-labels"),
    category_names="car,pedestrian,cyclist",
    images_rel_path="../images"
)

# Convert to YOLOv8 and save
YOLOv8ObjectDetectionOutput(
    output_file=Path("yolo-labels/data.yaml"),
    output_split="train"
).save(label_input=label_input)

Notes¶

KITTI format uses absolute pixel coordinates (x1,y1,x2,y2) for bounding boxes
Some fields like truncated, occluded, dimensions etc. are optional and can be set to -1 if unknown
The category name (type) should match one of the predefined categories when converting