KITTI detection example

Backgorund

KITTI detection dataset is used for 2D/3D object detection based on RGB/Lidar/Camera calibration data.

Data structure

When downloading the dataset, user can download only interested data and ignore other data. Each data has train and testing folders inside with additional folder that contains name of the data. So there are few ways that user can store this dataset:

1. Just unpack all archives in one folder

KITTI
|- image_2
    |- training
        |- image_2
            |- 000000.png
            |- 000001.png
            |- ...
    |- testing
        |- image_2
            |- 000000.png
            |- 000001.png
            |- ...
|- label_2
    |- training
        |- label_2
            |- 000000.txt
            |- 000001.txt
            |- ...
|- image_3
    |- training
        |- image_3
            |- 000000.png
            |- 000001.png
            |- ...
    |- testing
        |- image_3
            |- 000000.png
            |- 000001.png
            |- ...
|- calib
    |- training
        |- calib
            |- 000000.txt
            |- 000001.txt
            |- ...
    |- testing
        |- calib
            |- 000000.txt
            |- 000001.txt
            |- ...
|- velodyne
    |- training
        |- velodyne
            |- 000000.bin
            |- 000001.bin
            |- ...
    |- testing
        |- velodyne
            |- 000000.bin
            |- 000001.bin
            |- ...

1. Group them by training/testing

    |- training
        |- image_2
            |- 000000.png
            |- 000001.png
            |- ...
        |- image_3
            |- 000000.png
            |- 000001.png
            |- ...
        |- calib
            |- 000000.txt
            |- 000001.txt
            |- ...
        |- velodyne
            |- 000000.bin
            |- 000001.bin
            |- ...
        |- label_2
            |- 000000.txt
            |- 000001.txt
            |- ...            
    |- testing
        |- image_2
            |- 000000.png
            |- 000001.png
            |- ...
        |- image_3
            |- 000000.png
            |- 000001.png
            |- ...
        |- calib
            |- 000000.txt
            |- 000001.txt
            |- ...
        |- velodyne
            |- 000000.bin
            |- 000001.bin
            |- ...

Testing split has no label data

Data types

image_2/image_3: png images

calib: txt files with text and numbers

velodyne: bin files with lidar point clouds

label_2: txt files with class-name and box coordinates

API suggestion

We use second data storage format where data is in training/testing folders

import mldatasets as mlds

path = './KITTI'

def process_label(text):
    return process_text

def process_calib(text):
    return process_text

with mlds.seed(123):
    
    train, test = mlds.load(path) \
        .transform(["image_2", "image_3"], mlds.normalize) \
        .transform("calib", process_calib) \
        .transform("label_2", process_label) # prints "warning: mlds.transform('label2', process_label) ignored for 'testing' dataset

    train, val = train.split([0.81, -1])
    
    train_tf, test_tf, val_tf = mlds.all(train, test, val) \
        .shuffle() \
        .batch(32, remainder='cycle') \ # if datasize doesn't divide by 32, use first samples to pad (consider leaving batching to tensorflow or pytorch)
        .to_tensorflow()