Andrei Bursuc @CVPR#1697
馃敩 This weeks research highlight is 123D, KE:SAI's effort to unify all open driving data, creating the largest and most diverse pool of autonomous driving data out there.
Users are excited about the 123D platform unifying open driving datasets for KE:SAI because it enables training large-scale open foundation models without relying on proprietary sources.
馃摐 "123D: Unifying Multi-Modal Autonomous Driving Data at Scale"
https://arxiv.org/abs/2605.08084
I am particularly excited about 123D for KE:SAI because it enables us to train large-scale open foundation models without having to rely on proprietary data, making our work easily reproducible and easier to share.
Many big advancements in AI in recent years were preceded by a consolidation effort around data that enabled them.
You can find the open-source code on GitHub: https://github.com/kesai-labs/py123d
馃實 To solve this, KE:SAI has developed 123D, an open-source framework that unifies multimodal driving data through a single API.
Today, 123D already unifies 3300 hours of data spanning 90000 km of real-world driving from nuScenes, Waymo, Argoverse, and many others.
To name a few, Common Crawl enabled the training of large language models, LAION enabled the training of diffusion models for image generation, and Open X-Embodiment enabled robotics foundation models.
Now with 123D, you can release your data in a unified format that is compatible with all the existing datasets and directly benefit from new research breakthroughs.
We hope 123D will encourage more companies to contribute data to the open data ecosystem in the coming years.
Each dataset adopts different modalities: different cameras, lidars, ego states, annotations, HD maps, each with different rates and synchronization scheme.
It enables easily studying areas such as viewpoint robust 3D object detection or testing the generalization capabilities of reinforcement learning agents.
The 123D paper provides some baselines for these tasks, but there is still a lot of room for new methods to improve performance.
We hope the community leverages the 123D data to solve some of the important open generalization problems in autonomous driving.
馃殫 Autonomous driving has yet to see this type of consolidation.
Despite there being many different datasets available online, it is very hard to use them jointly.
馃彮 If you are a company that wants to spend the effort and time to open-source data, this used to present a significant risk.
Since your data will be incompatible with all the existing research dataset formats, your dataset might simply not get adopted by the research community.
If you try making your code compatible with all the different coordinate system conventions out there, you will quickly throw your PC out of the window.
123D is a collaboration between many institutions and people, in particular:
@DanielDauner, Valentin Charraut, @BastianBerle , Tianyu Li, Long Nguyen, Jiabao Wang, Changhui Jing, @MaxiIgl, Holger Caesar, @iamborisi , @yiyi_liao_, Andreas Geiger, and Kashyap Chitta.
Andrei Bursuc @CVPR#1697馃敩 This weeks research highlight is 123D, KE:SAI's effort to unify all open driving data, creating the largest and most diverse pool of autonomous driving data out there.