Founder

Think Autonomous

<h2>PointNet &amp; Voxel-Based 3D CNN from scratch</h2><p></p><p>End-to-End implementation of deep learning architectures for 3D point cloud understanding.</p><p></p><p>Built and trained <strong>PointNet</strong> and a <strong>Voxel-Based 3D CNN</strong> from scratch using the ShapeNet Core dataset — covering everything from data preprocessing to model visualization.</p><p></p><h2>🚀 PointNet Architecture (Point-Based Approach)</h2><p></p><p>Built the original <strong>PointNet</strong> architecture from scratch to classify and segment 3D shapes.</p><p></p><p><strong>Main Steps</strong></p><ul><li><p>ShapeNet Core dataset exploration</p></li><li><p>Custom Dataset and DataLoader</p></li><li><p>Model Architecture</p><p>T-Net — learns geometric transformations</p><p>FeatureNet — extracts global and local features</p><p>Classification / Segmentation Head — task-specific heads</p></li><li><p>Training and Evaluation</p></li><li><p>Inference and Visualization on individual 3D samples</p></li></ul><p></p><p><strong>Classification:</strong></p><div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmgxkeggguktn5thkc7fr3et8.png"><div class="image-overlay"></div></div></div></div><p><strong>Segmentation:</strong></p><div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmgxkf3b6ul5c5thk9fwm6jx5.png"><div class="image-overlay"></div></div></div></div><h2>🚀 Voxel-Based 3D CNN</h2><p></p><p>Built a voxelization pipeline to convert point clouds into voxel grids, then designed and trained a 3D CNN for classification.</p><p></p><p><strong>Main Steps</strong></p><ul><li><p>Dataset exploration on ShapeNet Core</p></li><li><p>Voxel Generation &amp; Visualization</p></li><li><p>Custom Dataset + DataLoader for voxel data</p></li><li><p>Model Architecture</p><p>Voxel Feature Encoding</p><p>3D CNN Feature Extraction</p><p>Classification Head</p></li><li><p>Training &amp; Validation Loop</p></li><li><p>Evaluation on test data with result visualization</p></li><li><p>Feature Visualization from intermediate CNN layers</p></li></ul><p></p><p><strong>Voxel Classification:</strong></p><div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmgxkgx28ufkz5tid2rv3ggbs.png"><div class="image-overlay"></div></div></div></div><p><strong>Feature Visualization:</strong></p><div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmgxkhf06ufsr5tid7gznemkc.gif"><div class="image-overlay"></div></div></div></div><p>GitHub repository: <a href="https://github.com/Rouizi/deep-point-cloud/tree/master" rel="noopener noreferrer nofollow" target="_blank">https://github.com/Rouizi/deep-point-cloud/tree/master</a></p>

End-to-End implementation of deep learning architectures for 3D point cloud understanding.

Deep Learning

Data Visualization

Deep Learning on Point Clouds

<p>3D perception pipeline and implementation inspired by SOTA methods.</p><p></p><p>Project Workflow:</p><p></p><ul><li><p>Run 3D object detector on sequences of LiDAR scans</p></li><li><p>Visualize results in both 2D and 3D</p></li><li><p>3D Association &amp; Tracking using:</p><p>Geometric cost</p><p>Appearance cost</p><p>3D IoU</p></li><li><p>Implement a 3D Kalman filter tracker based on SOTA research paper</p></li></ul><ul><li><p>Project tracked objects onto video frames</p></li><li><p>Visualize point clouds and 3D tracks in real-time</p></li></ul><p></p><p>GitHub repository: <a href="https://github.com/Rouizi/4d-perception" rel="noopener noreferrer nofollow" target="_blank">https://github.com/Rouizi/4d-perception</a></p><div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmgwgz2p8hgcc5thdc8qjcnp9.gif"><div class="image-overlay"></div></div></div></div>

In this project, I learned how to do 3D object detection, tracking, and visualization using LiDAR and camera data.

4D Perception: 3D Object Detection & Tracking

<p>Bird’s Eye View (BEV) transformation using both classical and deep learning methods.</p><p></p><p>It includes inverse perspective mapping and deep learning-based multi-camera fusion by implementing UNetXST.</p><p></p><p>Features:</p><ul><li><p>BEV generation using <strong>Inverse Perspective Mapping (IPM)</strong></p></li><li><p>Multi-camera to BEV transformation using <strong>deep learning approaches</strong></p></li><li><p>Implementation of the <strong>UNetXST architecture</strong> for BEV semantic segmentation</p></li><li><p>Training and evaluation on multiple front-facing camera inputs</p></li></ul><p></p><p>GitHub repository: <a href="https://github.com/Rouizi/bird-eye-view" rel="noopener noreferrer nofollow" target="_blank">https://github.com/Rouizi/bird-eye-view</a></p><p></p><div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmgwh5kf7hj9r5thd0mhjfx76.png"><div class="image-overlay"></div></div></div></div>

Bird’s Eye View (BEV) transformation using both classical and deep learning methods.

Machine Learning

Bird Eye View

<p>Early fusion and late fusion techniques to combine LiDAR point clouds and camera images for obstacle detection.</p><p></p><p>Projects:</p><p></p><p><strong>1. LiDAR-Camera Projection</strong></p><ul><li><p>Load and visualize Point Clouds</p></li><li><p>Load and visualize Images</p></li><li><p>Fuse LiDAR and Camera data manually</p></li><li><p>Project LiDAR Point Clouds onto Images</p></li></ul><p></p><div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmgwhbjuci03p5thkgra16mgr.gif"><div class="image-overlay"></div></div></div></div><p></p><p><strong>2. Early Fusion Project</strong></p><ul><li><p>Project 3D Point Clouds → 2D Images</p></li><li><p>Detect obstacles in 2D (Camera)</p></li><li><p>Fuse results for improved detection</p></li></ul><p></p><div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmgwhcloai0in5thk7t4of34f.gif"><div class="image-overlay"></div></div></div></div><p></p><p><strong>3. Late Fusion Project</strong></p><ul><li><p>Detect obstacles in 2D</p></li><li><p>Detect obstacles in 3D</p></li><li><p>Project 3D bounding boxes onto images</p></li><li><p>Fuse 3D LiDAR bboxes with 2D Camera bboxes</p></li></ul><p></p><p>GitHub repository: <a href="https://github.com/Rouizi/visual-fusion" rel="noopener noreferrer nofollow" target="_blank">https://github.com/Rouizi/visual-fusion</a></p>

Early fusion and late fusion techniques to combine LiDAR point clouds and camera images for obstacle detection.

Visual Fusion: Early & Late Fusion with LiDAR and Camera

Malt

GitHub

Upwork

3D Computer Vision Engineer | LiDAR, Point Clouds 

Skills: Python (Programming Language) · Computer Vision · Object Detection · Tensorflow · OpenCV · Deep Learning

Computer Vision Engineer

Skills: Machine Learning · Tensorflow · Python (Programming Language) · Image Processing · Deep Learning · Artificial Intelligence (AI) · Computer Vision · OpenCV · Object Detection

Skills: Web Development · Machine Learning · Tensorflow · Python (Programming Language) · Image Processing · Deep Learning · Computer Vision · Django · Object Detection

Yacine Rouizi

3D Computer Vision Engineer | LiDAR, Point Clouds

Recommended by

About me

Work experience

Computer Vision Engineer

Computer Vision Engineer

Freelance Blogger

Portfolio

Deep Learning on Point Clouds

Think Autonomous

4D Perception: 3D Object Detection & Tracking

Think Autonomous

Bird Eye View

Think Autonomous

Visual Fusion: Early & Late Fusion with LiDAR and Camera

Think Autonomous

Talents

Recruiters

About