ScanNet Dataset Files: A Community Request

Hello everyone!

I'm reaching out today with a request that I believe could benefit many in the 3D computer vision and machine learning community. Specifically, I'm looking to replicate results on the ScanNet dataset. You might know ScanNet as a fantastic dataset for scene understanding, reconstruction, and increasingly, for training and evaluating cutting-edge 3D generative models. However, I've encountered a common roadblock: accessing the dataset files themselves. The links previously available, particularly from resources like the OpenGaussian page, appear to be defunct. This has made it challenging for researchers like myself to get started or to reproduce existing work. I've even attempted to follow the preprocessing guidelines for ScanNet, but unfortunately, the resulting Gaussian splat renderings have been significantly blurred, indicating a potential issue with the data source or preprocessing pipeline that relies on accurate data.

To that end, I'm writing to inquire if it would be possible to share the files for the ScanNet dataset through a public, reliable link. Such a resource would be incredibly helpful for numerous individuals attempting to conduct research in this area. The ScanNet dataset is a cornerstone for many projects, and ensuring its accessibility is crucial for the continued progress and reproducibility in our field. Having a readily available and correctly preprocessed version of ScanNet would significantly lower the barrier to entry for new researchers and accelerate the pace of innovation. It would allow for more consistent benchmarking of new algorithms and facilitate the validation of existing research findings. I understand that hosting large datasets can be a challenge, but the community often rallies around such initiatives. Perhaps a shared drive, a torrent, or even updated links to official or archival sources could be established. I believe that fostering this kind of collaborative spirit is essential for advancing the state-of-the-art in 3D vision.

I've also taken the initiative to raise this issue in the OpenGaussian repository (link to issue), as this is a direct blocker for replicating their impressive results. The goal is not just to get the data for my own work, but to ensure that others who are inspired by groundbreaking research can follow suit without facing such fundamental obstacles. The reproducibility of scientific research is paramount, and dataset accessibility is a critical component of that. Without reliable access to foundational datasets like ScanNet, the field risks becoming fragmented, with research efforts becoming isolated and difficult to compare. The blurriness I've observed suggests that perhaps even if links were re-established, the data might require specific handling or a particular version to achieve the intended results. Therefore, any information or shared files that are known to be compatible with leading methods would be invaluable.

The importance of accessible datasets in machine learning cannot be overstated. Datasets are the bedrock upon which algorithms are built, trained, and tested. When a dataset like ScanNet, which has been instrumental in advancing scene understanding and 3D reconstruction, becomes difficult to access, it can stifle innovation. This is particularly true for generative models, where the quality and consistency of the training data directly impact the quality of the generated output. The blurriness I encountered is a stark reminder of how sensitive these models can be to data variations. Therefore, making ScanNet readily available, perhaps in a format that is known to work well with current state-of-the-art methods like Gaussian Splatting, would be a massive contribution to the research community. It would enable more robust comparisons, faster iteration on new ideas, and a more inclusive research environment.

In conclusion, my request is simple but significant: can we find a way to make the ScanNet dataset files publicly and reliably accessible? This would empower countless researchers, facilitate reproducibility, and foster further advancements in 3D computer vision. I'm hopeful that by bringing this to the community's attention, we can find a collaborative solution. If anyone has working links, knowledge of archival sources, or ideas on how to best share this crucial resource, please share them. Your contribution would be greatly appreciated and would undoubtedly help many in their research endeavors. Thank you for considering this request and for your dedication to advancing our field.

For further information on 3D computer vision and datasets, you might find these resources helpful:

• Stanford Large-Scale 3D Dataset (ScanNet): While direct download links can be tricky, searching academic archives or specific research group pages might yield results. Always verify the integrity and source of any downloaded data.
• Computer Vision Foundation: This organization is a key player in promoting computer vision research and often provides links to resources and datasets. You can explore their website for more information.
• arXiv: Many research papers related to ScanNet and 3D vision are published here. Searching for "ScanNet" on arXiv can lead you to papers that might mention data access or preprocessing steps.