DeepMill: Neural Accessibility Learning for Subtractive Manufacturing

DeepMill: Neural Accessibility Learning for Subtractive Manufacturing
Fanchao Zhong, Yang Wang, Peng-Shuai Wang, Lin Lu, and Haisen Zhao
ACM SIGGRAPH 2025 (conference track)

• DeepMill: Neural Accessibility Learning for Subtractive Manufacturing
  • 1. Environment Configuration
  • 2. Prepare Data
  • 3. Train
  • 4. Test
  • 5. Citation

1. Environment Configuration

(1) please install pytorch that fits your cuda version.

(2) install the packages required by DeepMill:

pip3 install -r requirements.txt

2. Prepare Data

To prepare the dataset, first store the accessibility results computed by traditional collision detection methods in the folder DeepMill/projects/data/raw_data. Assuming the subfolder is named models, place all accessibility data of the 3D models in DeepMill/projects/data/raw_data/models, where each .txt file corresponds to a single 3D model. Each line in the file contains 8 values representing a sampled point: 3D coordinates (X, Y, Z), normals (XN, YN, ZN), and two binary labels indicating whether the point is inaccessible or occluded. Additionally, the tool sizes used for each 3D model should be stored in DeepMill/projects/data/raw_data/models_cutter, with each .txt file containing four parameters that define the tool's shape. We provide part of the data we used in data/raw_data/data.zip. If you haven't installed Git LFS, please download data.zip(568M) separately by clicking on it in GitHub.

Once the data is in place, run the following commands to preprocess the dataset:

cd projects

python tools/seg_deepmill_cutter.py

If successful, the script will create two new folders, filelist and points, in DeepMill/projects/data. The dataset will be randomly split into a training set and a test set.

3. Train

Run the following command to train the network:

python run_seg_deepmill.py --depth 5 --model unet --alias unet_d5

Run the following command to visualize the training and testing process:

tensorboard --logdir logs

By default, test is performed every 50 epochs. If you do not need to output the predicted inaccessible points during each test (which can be time-consuming), you can comment out the following lines in projects/segmentation.py: self.visualization(batch['points'], logit_1, label, red_folder, gt_red_folder) self.visualization1(batch['points'], logit_2, label_2, green_folder, gt_green_folder)

A checkpoint of our model is provided in pretrained/00840solver.zip. Note that if you haven't installed Git LFS, please download 00840solver.zip(461M) separately by clicking on it in GitHub.

4. Test

Open the seg_deepmill.yaml file in projects/configs, and change run: train to run: test to perform testing only. In projects/data/vis, you can view the inaccessible points predicted by DeepMill for the 3D models in the test set.

5. Citation

If you find this project useful for your research, please kindly cite our paper:

@inproceedings{zhong2025deepmill,
  title={DeepMill: Neural Accessibility Learning for Subtractive Manufacturing},
  author={Zhong, Fanchao and Wang, Yang and Wang, Peng-Shuai and Lu, Lin and Zhao, Haisen},
  booktitle={Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers},
  pages={1--11},
  year={2025}
}

Feel free to drop us a message at [email protected] if you have any questions!