Awesome

Memory-based Adapters for Online 3D Scene Perception

Introduction

This repo contains PyTorch implementation for paper Memory-based Adapters for Online 3D Scene Perception based on MMDetection3D. Look here for 中文解读.

Memory-based Adapters for Online 3D Scene Perception
Xiuwei Xu*, Chong Xia*, Ziwei Wang, Linqing Zhao, Yueqi Duan, Jie Zhou, Jiwen Lu

We propose a model and task-agnostic plug-and-play module, which converts offline 3D scene perception models (receive reconstructed point clouds) to online perception models (receive streaming RGB-D videos).

News

• [2023/3/07]: Code released. Paper will be uploaded to Arxiv in next week.
• [2023/2/27]: Our paper is accepted by CVPR 2024.

Method

Overall pipeline of our work:

Memory-based adapters can be easily inserted into existing architecture by a few lines in config:

model = dict(
    type='SingleViewModel',
    img_memory=dict(type='MultilevelImgMemory', ...),
    memory=dict(type='MultilevelMemory', ...),
    ...)

Getting Started

For data preparation and environment setup:

• Installation
• Prepare Dataset

For training，evaluation and visualization:

• Train, Evaluation and Visualization

Main Results

We provide the checkpoints for quick reproduction of the results reported in the paper.

3D semantic segmentation on ScanNet and SceneNN:

3D object detection on ScanNet:

3D instance segmentation on ScanNet:

Visualization results:

Tips

If your GPU resources are limited, consider:

• Remove 2D modality (img_memory or the whole img_backbone). Note that in our 3D instance segmentation experiments, we remove img_memory to avoid OOM.
• Only insert adapters after high-level backbone features. We observe the higher the level, the better the performance of adapter, and the lower the resolution, the smaller the computation. For example, change:

img_memory=dict(type='MultilevelImgMemory', ada_layer=(0,1,2,3))
memory=dict(type='MultilevelMemory', vmp_layer=(0,1,2,3)),

        To:

img_memory=dict(type='MultilevelImgMemory', ada_layer=(2,3))
memory=dict(type='MultilevelMemory', vmp_layer=(2,3)),

        Then image and point cloud adapters will be only inserted after the highest two levels of features (for a four-level backbone).

Acknowledgement

We thank a lot for the flexible codebase of FCAF3D and valuable datasets provided by ScanNet and SceneNN.

Bibtex

If this work is helpful for your research, please consider citing the following BibTeX entry.

@article{xu2024online, 
      title={Memory-based Adapters for Online 3D Scene Perception}, 
      author={Xiuwei Xu and Chong Xia and Ziwei Wang and Linqing Zhao and Yueqi Duan and Jie Zhou and Jiwen Lu},
      journal={arXiv preprint arXiv:2403.06974},
      year={2024}
}