MCR-Lab

MCR-Lab is a pipeline for Manhole Center Regression evaluation of Deep Learning (via PyTorch) methods and classical approaches towards high-precision predictions in 3D LiDAR data (point clouds).

This should be only a light-weight pipeline using most likely HuggingFace and 3D utils for road extraction and circle fitting.

On Windows via Docker:

1. Start Docker Desktop and click on the console in the bottom-right
2. Navigate to the repo

   cd "D:\Studium\Master\Repos\MCR-Lab"

3. Build Docker (only once):

   docker build -f win.dev.Dockerfile -t mcrlab-img .

4. Run container:

   docker run --gpus all --rm -v .:/app mcrlab-img --config config/test.yaml
   # or
   docker run --gpus all -it --rm -v .:/app mcrlab-img --config config/test.yaml bash

5. Then connect with VSCode and the Docker Dev extension.
6. Run code via jupyter or with the docker console python ./src/main.py.

For training directly sart train.py?

On Windows via Anaconda:

1. DownloadAnaconda -> Anaconda Distribution, recommendation: one person installation (not all)
2. Install Anaconda via donwloaded installer -> default settings are most likely fine
3. Run Anaconda Prompt

   conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r
   conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/msys2

4. Install env

   conda create -n mcrlab python=3.12 pip -y
   conda activate mcrlab
   cd "C:\Users\tippolito\workspace\MCR-Lab"
   # you might want to install a specific PyTroch version before
   # https://pytorch.org/get-started/locally/
   pip install -e .

5. Start env (conda env list) + run python -> in anconda prompt

   conda activate mcrlab
   cd "C:\Users\tippolito\workspace\MCR-Lab"
   # cd "D:\Studium\Master\Repos\MCR-Lab" && D:
   # pip install -e .
   mcrlab --config "./configs/config.yaml"

Preparation-Phase:

• ✅ Make setup (with Docker)
• 📌 Add point-cloud loading utils
• Add point-cloud visualizing utils
• Inspect Paris-Lille-3D data (download it)
• Add Road Removing (which method is stable and easy?)
• Add BEV util + back-projection?
• Add Clustering util Cold-Practise-Phase:
• Add unsupervised geometry method (see master repo method idea) -> using implemented utils
• Try loading Mask2Former instance sementation and SAM model via HuggingFace
• Inspect results from methods (manually -> labels not available)
• Check and label real data

• Unsupervised Geometry Center Prediction
• Mask2Former-Instance-Segmentation (+ RANSAC shape-fitting to get center?)
• SAM-Segmentation + Classification-Head (manhole or not) (+ RANSAC shape-fitting to get center?)
• Point Transformer v3 (+ RANSAC shape-fitting to get center?) -> or direct regression?
  • maybe try https://huggingface.co/bryanchang/PTv3_laneline_segemenation_signal before you try use the official repo/pointcept