Teaching

MSIA - Point clouds

Evaluation:

Chose a paper and time slot here: LINK
A visio link will be added in the dedicated column.

Rules of the game

The idea is to do presentation in the style of a conference presentation.

  • 15 minutes slot
  • 7 minutes presentation of the paper with slides (better not be overtime)
  • 8 minutes discussion / questions on the paper and related notions.

WARNING - Colab notebooks

Copy / paste the link of the notebook (github link) in colab
Click on Copy to Drive oterwise you cannot save the changes to the notebook

Courses

From local properties to surface reconstruction

  • Local features
    • Normal estimation
    • Normal orientation
  • Surface reconstruction
    • Ball pivoting
    • Delaunay reconstruction
    • Poisson reconstruction
    • RANSAC
Slides Notebook Notebook (answers)

Descriptors and classic ML

  • Descriptors
    • Local descriptors
    • Global descriptors
    • Clustering
Slides Notebook Notebook (answers)

Images and graph

  • Image-based approaches
  • Graph-based approaches

Practical session

For the practical session, if not working on Google Colab directly, the material can be found on huggingface:


                from huggingface_hub import hf_hub_download
                hf_hub_download(repo_id="Msun/modelnet40", filename="modelnet40_ply_hdf5_2048.zip", repo_type="dataset", cache_dir=".")
                !unzip ./datasets--Msun--modelnet40/snapshots/d5dc795541800feeb7a4b3bd3142729a0d2adf7a/modelnet40_ply_hdf5_2048
Slides Notebook Notebook (answers)

From convolutions to transformers

  • Convolutions on points
  • Voxels
  • Mixers and transformers

Practical session

For the practical session, if not working on Google Colab directly, the material can be found on huggingface:


from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="wangps/shapenet_segmentation", filename="shapenetcore_partanno_segmentation_benchmark_v0_normal.zip", repo_type="dataset", cache_dir=".")
!unzip -qq ./datasets--wangps--shapenet_segmentation/snapshots/dbde146b974e1fc8628b47b1b1c4e50d8bc1a2ef/shapenetcore_partanno_segmentation_benchmark_v0_normal
!mv shapenetcore_partanno_segmentation_benchmark_v0_normal shape_data
Slides Notebook Notebook (answers)

Applications

  • Tasks
  • Self-supervised training
  • Domain adaptation
  • Open Vocabulary

import os
if not os.path.exists("./driving.hdf5"):
!wget https://github.com/aboulch/MSIA_points/releases/download/v0.0.0/driving.tar.gz
!tar -xvzf driving.tar.gz
Slides Notebook Notebook (answers)

Geometric vision

  • Structure from motion
  • Neural Radiance Fields (NeRFs)
  • Gaussian splatting
  • Dust3R and follow-ups
Slides Notebook Notebook (answers)