# BM-GCN

**Repository Path**: guohaolys/BM-GCN

## Basic Information

- **Project Name**: BM-GCN
- **Description**: Automatic coarse-to-fine method for cattle body measurement based on improved GCN and 3D parametric model
- **Primary Language**: Python
- **License**: MIT
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2025-02-19
- **Last Updated**: 2025-11-25

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# BM-GCN

This repository represents the source code for our paper "[Automatic coarse-to-fine method for cattle body measurement based on improved GCN and 3D parametric model](https://www.sciencedirect.com/science/article/pii/S0168169925001231?dgcid=author)".

<p align="center">
  <img src="assets/pipeline.png">
</p>

## Demo
<p align="center">
  <img src="assets/BM-GCN_gradio_demo.gif">
</p>

To start with, in command line, run the following to start the gradio user interface:
```
python app.py
```
Please refer to the GIF above for a step-by-step demonstration of the GUI usage.

## TODO List
- [x] Gradio based GUI.
- [x] Preprocessed data.
- [x] Training and inference. 
- [x] Evaluation code. 
- [ ] Online demo on HuggingFace Spaces.

## Installation
The code has been tested with CUDA 11.1, CuDNN 8.9, Python 3.9 and PyTorch 1.9.0; All experiments in our paper were conducted on a single NVIDIA GeForce 3070Ti GPU.
Our default installation method is based on Conda package and environment management:

### Set up conda environment
```
conda create -n bmgcn python=3.9 -y
conda activate bmgcn
```

### Install key packages
```
conda install pytorch==1.9.0 torchvision==0.10.0 pytorch-cuda=11.1 -c pytorch -c nvidia -y
conda install pytorch3d=0.7.2 -c pytorch3d -y
```

### Install pip dependency
```
pip install -r requirements.txt
```

## Data Preparation
Your folder structure should look as follows:
```bash
folder
├── checkpoints
│   └── ...
├── data
│   ├── cattle_coarse
│   ├── cattle_refine
│   ├── smal
│   ├── predefine
│   ├── priors
│   ├── mesh_down_sampling_4.npz
├── dataset
├── example_images
├── logs
├── model
├── smal_model
├── util
```

### Download Data
* Download the [cattle_coarse data](https://drive.google.com/file/d/1VtLs1hEEuX_TSi1JpErkDGZefhj3WQ-k/view?usp=sharing) and put it under the folder `./data/cattle_coarse`.

* Download the [cattle_refine data](https://drive.google.com/file/d/1VtLs1hEEuX_TSi1JpErkDGZefhj3WQ-k/view?usp=sharing) and put it under the foder `./data/cattle_refine`.

### Download SMAL and Priors
* Download the [SMAL](https://github.com/benjiebob/WLDO/tree/master/data/smal) template and the key point index from [index](https://drive.google.com/file/d/1icIjZGeSaVHkl7pMClfTMWKjI2AViS5s/view?usp=sharing) and put the downloaded files under `./data/smal`.

* Download the [predefined](https://drive.google.com/file/d/15Skn5-vlk4o6R-J96ggW4_knG0Pi3QtY/view?usp=sharing) body measurement index and put it under the folder `./data/predefine`.
  
* Download the [precomputed downsampling](https://drive.google.com/file/d/1gZhAnQOADLIEYT1apodChwJh3krfBQGB/view?usp=sharing) for the SMAL body mesh and put the downloaded file under `./data`.

* Download the [pose prior data](https://github.com/benjiebob/SMALify/tree/master/data/priors) and put the downloaded priors folder under `./data/priors`.

### Configurations
All configurations can be found in `./util/config.py`. If desired you can change the paths. 

## Train
### 1. Coarse Estimation Stage
To start with, in command line, run the following to start the first stage: 
```
python main_coarse.py 
```
### 2. Mesh Refinement Stage
Then you can continue to train stage 2 by running:
```
python main_refine.py --output_dir logs/refine --nEpochs 250 --lr 1e-5 --local_feat --gpu_ids 0 --feature_model pointnet2_msg_refine --save_results
```

## Test
We provide the [pretrained model](https://drive.google.com/file/d/1i9sFN8KEMpvjkZDxKnwQpIyJkkAmU9TB/view?usp=sharing). You can download and put it under the folder `./logs/refine/`. 
Test our model on the cattle data by running:
```
python eval.py --output_dir logs/test --resume logs/refine/model_best.pth.tar --gpu_ids 0 --local_feat --save_results
```
Qualitative results can be generated and saved by adding '--save_results' to the command. 

## Note
Currently, only `.pyc` files are provided. After setting up the environment, please replace all `.py` in the commands with `.pyc` to execute them. The full source code will be made publicly available upon the publication of the paper.

## Contact
For any questions on this project, please feel free to contact Hexiao Lu (luhexiao@cau.edu.cn).

## License
Please see [LICENSE](./LICENSE).