3D Object Recognition : Vision & Deep Learning

A two-part pipeline that goes from raw 3D meshes all the way to trained classifiers, combining computer vision (3D reconstruction from images) and deep learning (point cloud classification on ModelNet10/40).

---

Project structure

project/
├── vision/          # 3D reconstruction from Unity captures  →  README_VISION.md
└── deep_learning/   # Point cloud classification (PointNet, DGCNN…)  →  README_DL.md

---

Part 1 — Computer Vision : 3D Reconstruction

📄 See README_VIC.md for the full documentation.

Reconstructs 3D point clouds from synthetic images rendered in Unity, using four different methods of increasing complexity.

Tech stack : Unity 2021+, Python 3.x, OpenCV, NumPy, Pillow

Method	Technique	Status
A	Stereovision	Partial — disparity calibration issues
B	Canny + Epipolar geometry	Not working — insufficient keypoints
C	Multi-view depth maps (36 cameras)	Partial — coordinate frame mismatch Unity → OpenCV
D	Single-view depth map	Best results — partial reconstruction

Output formats : .ply (MeshLab visualisation) · .off (ModelNet10 compatible)

---

Part 2 — Deep Learning : Point Cloud Classification

📄 See README_DL.md for the full documentation.

Benchmarks five 3D classification architectures on ModelNet10 and ModelNet40, with a modular PyTorch codebase for easy model/sampler/dataset swapping.

Tech stack : Python 3.12, PyTorch, Open3D, TensorBoard, uv

Model	Paper	Key idea
SimplePointNet	—	Point-wise MLP + global max pool
PointNet	Qi et al., CVPR 2017	Learned spatial alignment (T-Net)
PointNet++	Qi et al., NeurIPS 2017	Hierarchical local feature learning
DGCNN	Wang et al., TOG 2019	Dynamic k-NN graph (EdgeConv)
PointTransformer	Zhao et al., ICCV 2021	Vector self-attention on local neighbourhoods

Supported datasets : ModelNet10 · ModelNet40

---

End-to-end pipeline

Unity scene
    │
    │  Capture (RGB images + depth maps)
    ▼
3D Reconstruction  (vision/)
    │
    │  .off files
    ▼
Point Cloud Classification  (deep_learning/)
    │
    ▼
Predicted class  (chair, table, bathtub…)

The .off files produced by the vision pipeline are directly compatible with the deep learning pipeline's dataset format.

---

Quick links

Resource	Link
ModelNet10	Kaggle·Princeton
ModelNet40	Kaggle
Unity data (Google Drive)	Drive folder
PointNet	https://arxiv.org/abs/1612.00593
PointNet++	https://arxiv.org/abs/1706.02413
DGCNN	https://arxiv.org/abs/1801.07829
Point Transformer	https://arxiv.org/abs/2012.09164