# CALoss

**Repository Path**: csc105/CALoss

## Basic Information

- **Project Name**: CALoss
- **Description**: The codes for Learning to Measure the Point Cloud Reconstruction Loss in a Representation Space
- **Primary Language**: Python
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2023-03-07
- **Last Updated**: 2024-02-29

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# CALoss
[CVPR'23] The codes for Learning to Measure the Point Cloud Reconstruction Loss in a Representation Space

## Environment
* TensorFlow 1.13.1
* Cuda 10.0
* Python 3.6.9
* numpy 1.14.5

We also provide an available conda environment (`lcd.yaml`) in this repo. Please run:

```
conda env create -f caloss.yaml
```

## Dataset
The adopted ShapeNet Part dataset is adopted following [FoldingNet](http://www.merl.com/research/license#FoldingNet), while the ModelNet10 and ModelNet40 datasets follow [PointNet](https://github.com/charlesq34/pointnet.git). Other datasets can also be used. Just revise the path by the (`--filepath`) parameter when training or evaluating the networks.
The files in (`--filepath`) should be organized as

        <filepath>
        ├── <trainfile1>.h5 
        ├── <trainfile2>.h5
        ├── ...
        ├── train_files.txt
        └── test_files.txt

where the contents in (`train_files.txt`) or (`test_files.txt`) should include the directory of training or testing h5 files, such as:

        train_files.txt
        ├── <trainfile1>.h5
        ├── <trainfile2>.h5
        ├── ...

We also provide the processed datasets in [Google Drive](https://drive.google.com/file/d/1sjUk8o-wsZp2PJUej4TsmjnOPvjJegKR/view?usp=sharing). You can just download, unzip them, and just set filepath to one of the dataset paths.

## Usage

1. Preparation

```
cd ./tf_ops
bash compile.sh
```

2. Train

For the reconstruction task,
```
Python3 vv_cae.py
```

Note that the paths of data should be edited through the (`--filepath`) parameter according to your setting. For example, if we use the download dataset (`./objdata/ShapeNet_part`), the training command would be

```
Python3 vv_cae.py --filepath ./objdata/ShapeNet_part
```

3. Test

For the evaluation of reconstruction errors,
```
Python3 vvae_eva.py 
```

The trained weight files should be provided by the (`--savepath`) parameter to evaluate the performances.

Here, we also provide [weights](https://drive.google.com/file/d/19IqJ-LV5zpVrstu2yCn-LcUnt9mLaPtq/view?usp=sharing) of the reconstruction network AE pre-trained on ShapeNet Part dataset. To evaluate its performance, just download and unzip it, then change the savepath to its folder.
If the dataset and weights are put in (`./objdata/ShapeNet_part`) and (`./pnae`), respectively, the command would be

```
Python3 vvae_eva.py --filepath ./objdata/ShapeNet_part --savepath ./pnae
```

4.复现信息

复现人：钱毅杰
复现时间：2024.1.31
复现小结：本文提出了一种新的点云距离度量方法，使用网络将三维点云映射到高维的表征空间，并使用表征距离度量点云之间的形状差异。在训练过程中，该映射网络同时使用对比学习和对抗策略进行动态的更新，以构建相似形状表征较近但重构缺陷比较突出的表征空间。本方法的缺点在于，由于引入了网络结构，训练所需的显存会增加，且需要调整的超参数也有相应的增加。