6Img-to-3D: Few-Image Large-Scale Outdoor Driving Scene Reconstruction
Theo Gieruc*¹², Marius Kaestingschaefer*¹, Sebastian Bernhard¹, Mathieu Salzmann²,

¹Continental AG, ²EPFL *denotes equal contribution

A PyTorch implementation of the 6Img-to-3D model for large-scale outdoor driving scene reconstruction. The model takes as input six images from a driving scene and outputs a parameterized triplane from which novel views can be rendered.

If you find this code useful, please reference in your paper:

@misc{gieruc20246imgto3d,
      title={6Img-to-3D: Few-Image Large-Scale Outdoor Driving Scene Reconstruction}, 
      author={Théo Gieruc and Marius Kästingschäfer and Sebastian Bernhard and Mathieu Salzmann},
      year={2024},
      eprint={2404.12378},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Inward and outward-facing camera setups differ significantly in their view overlap. Outward-facing (inside-out) camera setups overlap minimally, whereas inward-facing (outside-in) setups can overlap across multiple cameras.

Given six input images, we first encode them into feature maps using a pre-trained ResNet and an FPN. The scene coordinates are contracted to fit the unbounded scenes. MLPs, cross-and self-attention layers form the Image-to-Triplane Encoder of our framework. Images can be rendered from the resulting triplane using our renderer. We additionally condition the rendering process on projected image features.

conda create -n sixtothree python=3.8
conda activate sixtothree

Install PyTorch 2.0.1 with CUDA 11.8 (recommanded), cuda-toolkit and tinycudann.

pip install torch==2.0.1+cu118 torchvision==0.15.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118 
conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit
pip install ninja git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

Install the MMLAB suite

pip install mmdet==2.20.0 mmengine==0.8.4 mmsegmentation==0.20.0  mmcls==0.25.0 mmcv-full==1.5.0 -f https://download.openmmlab.com/mmcv/dist/cu118/torch2.0.1/index.html

Install the other dependencies

pip install tensorboardX crc32c pandas pyyaml==6.0.1  imageio==2.33.1 imageio-ffmpeg==0.4.9 lpips==0.1.4 pytorch-msssim==1.0.0 kornia==0.7.0 yapf==0.40.1 jupyter notebook seaborn==0.13.2

To train the model use the train.py script

So that the training runs faster, we first turn the train dataset into pickles with the pickles_generator.py script.

 python utils/generator_pickles.py --dataset-config config/_base_/dataset.py --py-config config/config.py

To train the model, run the training script with the desired arguments specified using the command line interface. Here's how to use each argument:

• --py-config: Path to the Python configuration file (.py) containing model configurations. This file specifies the architecture and parameters of the model being trained.

• --ckpt-path: Path to a TPVFormer checkpoint file to initialize model weights from. If specified, the training will resume from this checkpoint.

• --resume-from: Path to a checkpoint file from which to resume training. This option allows you to continue training from a specific checkpoint.

• --log-dir: Directory where Tensorboard training logs and saved models will be stored. If not provided, logs will be saved in the default directory with a timestamp.

• --num-scenes: Specifies the number of scenes to train on. This argument allows for faster training when only a subset of scenes is required.

• --from-epoch: Specifies the starting epoch for training. If training is interrupted and resumed, you can specify the epoch from which to resume training.

To run the train9ing script, execute the Python file train.py with the desired arguments specified using the command line interface. For example:

python train.py --py-config config/config.py --ckpt-path ckpts/tpvformer.pth --log-dir evaluation_results 

To evaluate the model, use the eval.py script.

The evaluation script can be run with different options to customize the evaluation process. Here's how to use each argument:

• --py-config: Path to the Python configuration file (.py) containing model configurations. This file specifies the architecture and parameters of the model being evaluated.

• --dataset-config: Path to the dataset configuration file (.py) containing dataset parameters. This file specifies dataset-specific settings such as image paths and scalling.

• --resume-from: Path to the checkpoint file from which to resume model evaluation. This argument allows you to continue evaluation from a previously saved checkpoint.

• --log-dir: Directory where evaluation Tensorboard logs and results will be saved. The default behavior is to create a directory with a timestamp indicating the evaluation start time.

• --depth: If specified, depth maps will also be saved.

• --gif: If specified, the script generates GIFs from the evaluated images.

• --gif-gt: If specified, GIFs are generated for ground truth images.

• --img-gt: If specified, the script saves ground truth images alongside the generated images.

• --num-img: Specifies the number of images to evaluate. By default, all images in the dataset are evaluated. This argument allows for faster evaluation when only a subset of images is required.

• --time: Compute inference time of the model and save results in t_decode.txt, t_encode.txt

To run the evaluation script, execute the Python file eval.py with the desired arguments specified using the command line interface. For example:

python eval.py --py-config ckpts/6Img-to-3D/config.py --resume-from ckpts/6Img-to-3D/model_checkpoint.pth --log-dir evaluation_results --depth --img-gt --dataset-config config/_base_/dataset_eval.py

Copyright (C) 2024 co-pace GmbH (subsidiary of Continental AG). All rights reserved. This repository is licensed under the BSD-3-Clause license. See LICENSE for the full license text.