This dataset is created by MonoTTA: Fully Test-Time Adaptation for Monocular 3D Object Detection, based on KITTI. You can check this link for more details: https://arxiv.org/abs/2405.19682v1 And access the code: https://github.com/Hongbin98/MonoTTA Please double-check the demands of KITTI when you try to download this dataset and obey their rules.

Context

Public dataset for KITTI Object Detection: https://github.com/DataWorkshop-Foundation/poznan-project02-car-model

https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1246155%2Fc9cc7e9e46ce68919b8157f82b4c0d06%2Fpassat_sensors_920.png?generation=1605764967434311&alt=media" alt="">

Licence

Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License

When using this dataset in your research, we will be happy if you cite us: @INPROCEEDINGS{Geiger2012CVPR, author = {Andreas Geiger and Philip Lenz and Raquel Urtasun}, title = {Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite}, booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)}, year = {2012} }

Introduction

The dataset was built in 2020 as part of the thesis project titled "Real-Time Object Detection with Deep Learning on an Embedded GPU System" by Márk Antal Csizmadia that was submitted in partial fulfillment of the requirements for the degree of Bachelor of Engineering in Electronic Engineering at the University of Manchester, UK. The dataset is part of the public domain.

The annotated objects in the dataset include six-sided boardgame dices (dice), AAA, AA, and 9 V batteries (battery), toy cars (toycar), spoons (spoon), highlighters (highlighter), and tea candles (candle). The dataset was built through different means that included scraping images off the Internet with the Bing Image Search API, remixing existing datasets from the public domain, extracting video frames from videos downloaded from YouTube in line with its fair-use policy, and manually taking photographs.

There are in overall 1644 images in the dataset that contain 2815 objects. The distribution of the objects in the dataset are as shown in the table below.

The images were resized into 640 x 640 pixels and were padded to keep the original aspect ratio. The resized images were annotated using an annotation tool published in the public domain. The annotation of the full dataset took around 5 weeks. This, unfortunately, should have been done as a pre-processing step before training an algorithm, but at the time when I built this dataset, I was not yet aware of that.

The specific labeling tool was selected since it produces annotation data in the KITTI format. The format defines a set of parameters for each object in each image that includes type, truncated, occluded, alpha, bbox, dimensions, location, rotation_y, and score. The type parameter describes the object type which can be one of “dice”, “toycar”, “battery”, “candle”, "spoon", and "highlighter". The bbox parameter is an ordered set of four coordinates that define the top-left, and the bottom-right vertices of the ground-truth bounding box. The rest of the parameters are further described in the original source.

Unfortunately, there are some missing annotations of the objects of interest, such as that in 00000331.jpg. This issue is not significant and does allow to train accurate object detection models.

Complex-yolov5 for 3d object detection. This is an unofficial edition of Complex-YOLO merged with yolov5s, together with newly added visualization for pointcloud in ./src/test.py. Dataset is available at: https://www.kaggle.com/datasets/dingdangar/kitti-3d-complexyolo. Complex-YOLO: https://github.com/maudzung/Complex-YOLOv4-Pytorch YOLOv5: https://github.com/ultralytics/yolov5 There is a pretrained model (77 epoches) in ./checkpoints/complexer_yolo/