Hi, and thanks for open-sourcing LiDARCrafter.
I have been trying to reproduce the Layout2Scene and Scene2Seq parts of the pipeline, and I wanted to check whether I am understanding the released checkpoints and intended training setup correctly.
-
From the public Hugging Face repo, I was able to find the evaluation backbones under pretrained_models/evaluation, as well as an object checkpoint such as pretrained_models/training/nuscenes-object-1000000.pth. However, I could not find public pretrained checkpoints for Layout2Scene or for the Scene2Seq autoregressive temporal model.
-
I also noticed that the code references filenames like nuscenes-lox-layout-v4-500000.pth and nusc-auto-reg-v2-350000.pth, but I could not find those files in the public release.
-
Because of that, I am currently training the range-image stage from scratch. At the moment, my understanding is that pretrained_models/evaluation mainly contains evaluator, detector, or feature-extractor weights rather than the generator checkpoints, and that the pretrained checkpoints for Layout2Scene and Scene2Seq may not have been publicly released. Based on that, I have been assuming that training these stages from scratch is the expected workflow for reproducing the paper.
Could you let me know whether that understanding is correct?
I would also really appreciate any guidance on which released config is closest to the paper’s Layout2Scene setup.
- In particular, I was unsure whether
nuscenes-box-layout-v6 is intentionally a “without ground” variant, and whether the paper’s figures or results correspond more closely to nuscenes-box-layout-v5 or to another config. If there is any internal or unofficial mapping between the paper stages and the config names or checkpoint names used in the codebase, that would also be very helpful.
Thanks again for making the project public. Any clarification would be greatly appreciated.
Hi, and thanks for open-sourcing LiDARCrafter.
I have been trying to reproduce the
Layout2SceneandScene2Seqparts of the pipeline, and I wanted to check whether I am understanding the released checkpoints and intended training setup correctly.From the public Hugging Face repo, I was able to find the evaluation backbones under
pretrained_models/evaluation, as well as an object checkpoint such aspretrained_models/training/nuscenes-object-1000000.pth. However, I could not find public pretrained checkpoints forLayout2Sceneor for theScene2Seqautoregressive temporal model.I also noticed that the code references filenames like
nuscenes-lox-layout-v4-500000.pthandnusc-auto-reg-v2-350000.pth, but I could not find those files in the public release.Because of that, I am currently training the range-image stage from scratch. At the moment, my understanding is that
pretrained_models/evaluationmainly contains evaluator, detector, or feature-extractor weights rather than the generator checkpoints, and that the pretrained checkpoints forLayout2SceneandScene2Seqmay not have been publicly released. Based on that, I have been assuming that training these stages from scratch is the expected workflow for reproducing the paper.Could you let me know whether that understanding is correct?
I would also really appreciate any guidance on which released config is closest to the paper’s
Layout2Scenesetup.nuscenes-box-layout-v6is intentionally a “without ground” variant, and whether the paper’s figures or results correspond more closely tonuscenes-box-layout-v5or to another config. If there is any internal or unofficial mapping between the paper stages and the config names or checkpoint names used in the codebase, that would also be very helpful.Thanks again for making the project public. Any clarification would be greatly appreciated.