Home

Awesome

Fine-grained Image-to-LiDAR Contrastive Distillation with Visual Foundation Models (NeurIPS2024)

Official PyTorch implementation of the method OLIVINE. More details can be found in the paper:

Fine-grained Image-to-LiDAR Contrastive Distillation with Visual Foundation Models, NeurIPS2024 [arXiv] by Yifan Zhang and Junhui Hou.

Overview of the method

Dependencies

Please install the required required packages. Some libraries used in this project, including MinkowskiEngine and Pytorch-lightning are known to have a different behavior when using a different version; please use the exact versions specified in requirements.txt.

Datasets

The code provided is compatible with nuScenes and semantic KITTI. Put the datasets you intend to use in the "datasets" folder (a symbolic link is accepted).

datasets/
├── nuscenes
    ├── camseg (semantic labels infered by Grounded-SAM)
    ├── lidarseg (decompress nuScenes-lidarseg-all-v1.0.tar)
    ├── maps
    ├── samples
    ├── sweeps
    ├── v1.0-mini
    ├── v1.0-test
    ├── v1.0-trainval
    └── zip_files
└── semantic_kitti
    ├── dataset
        ├── poses
        └── sequences
└── other datasets...
<!-- ## Pre-trained models Minkowski SR-UNet pre-trained on nuScenes: [Link](https://drive.google.com/file/d/1gMzwFymKkSi5WqeEL8Ie9L8N6qTcli2u/view?usp=sharing). --> <!-- ### SPconv VoxelNet [VoxelNet pre-trained on nuScenes](#) (to release) [PV-RCNN finetuned on KITTI](#) (to release) -->

Reproducing the results

Predict the weak semantic labels (required)

<!-- First we use the [Grounded-SAM](https://github.com/IDEA-Research/Grounded-Segment-Anything) to obtain weak semantic labels of RGB images. To initialize the submodle: ``` git submodule update --init ``` Then, the script for the prediction can be found in **Grounded-SAM/infer.sh**. Please install Grounded-SAM following the instructions (see **Grounded-SAM/README.md**) before running the script. You can also obtain the labels by directly downloading the files we provide in [Baidu netdisk](https://pan.baidu.com/s/1CNQRCawQuL5eNFQqTXMwlw?pwd=1jkr) or [Google Drive](https://drive.google.com/file/d/1FGow9PFoH11M-g_eUg-7ZGZdT_8pUt4j/view?usp=drive_link). -->

First we use the SEEM to obtain weak semantic labels of RGB images. If you do not want to execute the following steps by yourself, you can also obtain the labels by directly downloading the files we provide in Baidu netdisk or Google Drive.

  1. Install necessary libraries in demo_code/requirements.txt
  2. Link nuScenes dataset to demo_code/data/sets. Command: ln -s datasets/nuscenes demo_code/data/sets/
  3. Go to demo_code directory and run the script bash semantic_label_generation.sh
  4. Organize the generated files and put them in data/nuscenes/camseg.

Pre-training a 3D backbone

To launch a pre-training of the Minkowski SR-UNet (minkunet) on nuScenes:

python pretrain.py --cfg config/olivine_minkunet.yaml

You can alternatively replace minkunet with voxelnet to pre-train a PV-RCNN backbone.
Weights of the pre-training can be found in the output folder, and can be re-used during a downstream task. If you wish to use multiple GPUs, please scale the learning rate and batch size accordingly.

TIPs: The pre-trained weights in the final epoch of pre-training may not always be the best; it's worth considering saving the weights from other rounds, such as the 40th epoch.

Semantic segmentation

To launch a semantic segmentation, use the following command:

python downstream.py --cfg_file="config/semseg_nuscenes.yaml" --pretraining_path="output/pretrain/[...]/model.pt"

with the previously obtained weights, and any config file. The default config will perform a finetuning on 1% of nuScenes' training set, with the learning rates optimized for the provided pre-training.

To re-evaluate the score of any downstream network, run:

python evaluate.py --resume_path="output/downstream/[...]/model.pt" --dataset="nuscenes"

If you wish to reevaluate the linear probing, the experiments in the paper were obtained with lr=0.05, lr_head=null and freeze_layers=True.

Object detection

All experiments for object detection have been done using OpenPCDet.

Published results

All results are obtained with weights pre-trained on nuScenes.

Few-shot semantic segmentation

Results on the validation set using Minkowski SR-Unet:

MethodnuScenes<br />lin. probingnuScenes<br />Finetuning with 1% dataKITTI<br />Finetuning with 1% data
Random init.8.130.339.5
PointContrast21.932.541.1
DepthContrast22.131.741.5
PPKT36.437.843.9
SLidR38.838.344.6
OLIVINE50.050.549.3

Semantic Segmentation on nuScenes

Results on the validation set using Minkowski SR-Unet with a fraction of the training labels:

Method1%5%10%25%100%
Random init.30.347.756.664.874.2
SLidR39.052.258.866.274.6
OLIVINE50.660.265.070.176.5

Object detection on KITTI

<!-- #### Results on the validation set using Minkowski SR-Unet with a fraction of the training labels: Method |5% |10% |20% --- |:-: |:-: |:-: Random init. |56.1 |59.1 |61.6 [PPKT](https://arxiv.org/abs/2104.04687)|**57.8**|60.1 |61.2 SLidR |**57.8**|**61.4**|**62.4** --> <!-- ## Unpublished preliminary results -->

All results are obtained with a pre-training on nuScenes.

Results on the validation set using PV-RCNN:

MethodCarPedestrianCyclistmAP@40
Random init.84.557.971.371.3
STRL*84.757.871.971.5
PPKT83.255.573.870.8
SLidR84.457.374.271.9
OLIVINE84.859.374.272.8

*STRL has been pre-trained on KITTI, while SLidR and PPKT were pre-trained on nuScenes

Results on the validation set using SECOND:

MethodCarPedestrianCyclistmAP@40
Random init.81.550.966.566.3
DeepCluster*66.1
SLidR81.951.668.567.3
OLIVINE82.053.269.868.3

*As reimplemented in ONCE

<!-- ## Visualizations For visualization you need a pre-training containing both 2D & 3D models. We provide the raw [SR-UNet & ResNet50 pre-trained on nuScenes](https://github.com/valeoai/SLidR/releases/download/v1.1/minkunet_slidr_1gpu_raw.pt). The image part of the pre-trained weights are identical for almost all layers to those of [MoCov2](https://github.com/facebookresearch/moco) (He et al.) The [visualization code](utils/visualization.ipynb) allows to assess the similarities between points and pixels, as shown in the article. -->

Acknowledgment

We implement the method based on SLidR. Part of the codebase has been adapted from PointContrast. Computation of the lovasz loss used in semantic segmentation follows the code of PolarNet.

License

OLIVINE is released under the Apache 2.0 license.

Citation

If you use OLIVINE useful in your research, please consider citing:

@inproceedings{zhang2024fine,
  title={Fine-grained Image-to-LiDAR Contrastive Distillation with Visual Foundation Models},
  author={Zhang, Yifan and Hou, Junhui},
  booktitle={Advances in Neural Information Processing Systems},
  year={2024}
}