All classes Occlusion Robust Face Recognition Syllabus · Episode 4 of 6

Face Occlusion Robust Recognition Introduction #4: Multi-Scale Feature Fusion & Face Embedding with ArcFace Loss

0 views

ArcFace loss and multi-scale feature fusion (FPN) tutorial with PyTorch: 128-dimensional face embedding, additive angular margin, occlusion augmentation. Episode 4 of the Face Occlusion Robust series.

Software: Deep Learning (PyTorch ArcFace)
Level: Advanced
Duration: ± 16 min
Series total: 1 h 26 min · 6 videos
Published: October 24, 2026

Episode 4 of 6 · view syllabus

Subscribe ▶ Watch the video lesson0 plays

Checking access...

Want to go deeper?

Join one of our courses or ask us directly on WhatsApp.

View the full 6-episode syllabus View Face Recognition Courses Ask on WhatsApp

Recommended classes for you

Picked from the software, topic, and level of “Face Occlusion Robust Recognition Introduction #4: Multi-Scale Feature Fusion & Face Embedding with ArcFace Loss”.

6 episodesFree preview

Python · Intermediate–Advanced

Deepfake Detection Vision Transformer

6 video lessons with PDF guides, 1 h 11 min in total. Starts with: Deepfake Concepts, Manipulation Types & Real vs Fake Face Dataset.

1 h 11 min 0 0

View class

NSGA-II MOPSO Hybrid Multi-Objective Optimization

6 episodesFree preview

Python · Intermediate–Advanced

NSGA-II MOPSO Hybrid Multi-Objective Optimization

6 video lessons with PDF guides, 59 min in total. Starts with: Multi-Objective Optimization: Pareto Front, Dominance, and Test Functions ZDT/DTLZ.

59 min 9 0

View class

18 episodesFree preview

Python · Beginner–Advanced

Machine Learning Python

18 video lessons with PDF guides, 3 h 41 min in total. Starts with: Introduction to Machine Learning & First Model.

3 h 41 min 61 0

View class

100 episodesFree preview

Claude · Beginner–Advanced

Claude from Zero to Expert

100 video lessons with PDF guides, 15 h 1 min in total. Starts with: What Claude Is and How It Differs from ChatGPT and Gemini.

15 h 1 min 62 0

View class

50 episodesFree preview

Agentic AI · Beginner–Intermediate

Agentic AI Without Coding

50 video lessons with PDF guides, 14 h 29 min in total. Starts with: Chatbots, Automation, and AI Agents: What's the Difference?.

14 h 29 min 66 0

View class

6 episodesFree preview

Python · Intermediate–Advanced

Brain Tumor Segmentation with U-Net

6 video lessons with PDF guides, 1 h 10 min in total. Starts with: Brain MRI Dataset & Tumor Mask: Loading, Visualization & Patient-wise Splitting.

1 h 10 min 28 0

View class

Deep Learning CNN-LSTM Pneumonia Detection

6 episodesFree preview

Python · Intermediate–Advanced

Deep Learning CNN-LSTM Pneumonia Detection

6 video lessons with PDF guides, 57 min in total. Starts with: Chest X-ray Dataset: Loading, Exploration & Data Splitting.

57 min 22 0

View class

6 episodesFree preview

Python · Intermediate–Advanced

Transfer Learning for Leaf Disease

6 video lessons with PDF guides, 1 h 5 min in total. Starts with: PlantVillage Dataset: Folder Structure, Class Distribution & Data Split.

1 h 5 min 9 0

View class

Need more than a video?

Other ways we can help with Python.

Learn with a mentor

Scheduled AI & Machine Learning courses with live guidance and a certificate.

Course

Multi-Objective Optimization Course: NSGA-II, MOPSO & Hybrid Algorithms (Python)

Course

ECG Arrhythmia Classification Course with Wavelet & Deep Learning (PyTorch)

Course

ANFIS + PSO/GA Course for Non-Linear Time Series Forecasting (Python)

See all in AI & Machine Learning →

Have it built for you

No time to learn Python? Our team can do the project for you.

Service

Python Machine Learning — Prediction Models, Customer Segmentation, Fraud Detection & AI Apps

Service

AI Image Creation with Stable Diffusion — Unlimited Custom Visual Production

★ 4.9

Service

AI Image & Video Creation with Leonardo AI — All-in-One Visuals for Business

★ 4.9

See all in AI & Machine Learning →

Get notified about new Python classes

At most one email a week. Unfollow anytime with one click.

Transcript

The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.

Show transcript

Introduction

Here is the final result, each face is transformed into 128 numbers, and faces of the same person are placed close to each other. We are training three models with R-Face Loss, and one of the comparisons will be very decisive for the occlusion resistance in the final episode. Let's start. Hello, welcome to episode four of the occlusion-resistant face recognition series. Modern face recognition does not classify faces into a fixed list of names, but rather transforms faces into embedding factors. Two faces are compared by the angle between their vectors, so new people can be registered without retraining. Today we are combining multi-scale features into an embedding and training it with R-Face Loss.

Step 1: Multi-scale fusion & single-scale comparison

Create a model_underscore_face.py file. The multi-scale fusion class uses FPN from the previous episode. Each feature map P2 to P5 is globally averaged into a 128 factor, the four are combined into 512 numbers, then projected with a linear layer and B-Norm into a 12,8-dimensional normalized embedding. As a comparison, the single-class class only uses the last layer feature of ResNet 18. Now R-Face Loss.

Step 2: ArcFace loss (additive angular margin)

Each person has a weight factor that represents the center of their class. The cosine between the embedding and each center is calculated. Then, for the correct class, the angle is added with marine M of 0.5 radians before taking the cosine again. As a result, the model is forced to make the embedding much closer to its own class center than what softmax normally requires. All logits are scaled by S30, then entered into cross entropy. This is the additive angular marine that makes R-Face highly discriminative.

Step 3: Occlusion augmentation & identification accuracy

Create a train_underscore_face.py file. We load 112-pixel faces and use GPU if available. The augmentation function performs horizontal flipping, and with a 50% chance, adds one random occlusion, mask, glasses, or hand, placed using a landmark template because the face is already aligned. The accuracy identification function calculates the gallery embedding from the training data, averages it per person into a centroid, then each validation face is recognized as the centroid with the highest cosine.

Step 4: Training three models

The train function uses Adamwe, Cosine Anialing, and 12 Ipok. The occlusion parameter determines whether occlusion augmentation is used. The best model based on validation accuracy is saved. We will train three models: multi-scale fusion with occlusion augmentation, single-scale with occlusion augmentation, and multi-scale fusion without occlusion augmentation. Run.

Step 5: Training results

Each epoch only takes a few seconds on the GPU. The Arc Loss Phase starts high, in the tens, due to the large marine, then drops quickly. All three models achieve validation identification accuracy around 95-97%. On faces without occlusion, all three are almost the same, and multi-scale fusion is not much different from single-scale. The real difference becomes visible when the face is covered, which we will test in the final episode. Done.

Done!

We have built multi-scale feature fusion, 128-dimensional face embedding, ArcFace loss with additive angular marine, and random occlusion augmentation, then trained three comparison models. In the next episode, we will use this embedding for face verification, cosine similarity, threshold, ROC curve, and IIR. Currently working on a thesis, dissertation, or AI research in the field of face and biometric recognition? Cerdas Inovasi Digital helps implement deep learning, from synthetic occlusion, face alignment, feature pyramid, ArcFace, verification, and ROC, up to face attendance applications, and opens deep learning courses with PyTorch. Want to master the entire series, from face datasets to attendance applications? All 6 episodes, complete with PDF step-by-step guides and source code, are available to premium tutorial subscribers. Subscribe starting at Rp69.000 per month, save 28% with the 12-month package, or pay once Rp999.000 for lifetime access. Want just one class? Buy the single class with lifetime access. Subscribe now at the premium page cerdasinovasidigital.com Contact us via WhatsApp, or visit cerdasinovasidigital.com Thank you for watching, see you in the next episode.

Ask the class material

Type a question or keyword; we find the part of this episode that explains it and jump the video there.

My notes

Notes are saved with the current video time and are only visible to you.

Questions & answers

Stuck on a step? Ask here; the instructor and other learners can answer.

No questions yet.

Your instructor

Cerdas Inovasi Digital Team

Software developers, BIM & 3D engineers, and AI practitioners

53 video classes1,296 episodes

We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.

See our clients →

Memuat

Recommended classes for you

Need more than a video?

Transcript

0:03Introduction

0:49Step 1: Multi-scale fusion & single-scale comparison

3:56Step 2: ArcFace loss (additive angular margin)

5:13Step 3: Occlusion augmentation & identification accuracy

8:29Step 4: Training three models

13:58Step 5: Training results

14:31Done!

Ask the class material

My notes

Questions & answers

Your instructor

Recommended classes for you

Need more than a video?

Transcript

0:03Introduction

0:49Step 1: Multi-scale fusion & single-scale comparison

3:56Step 2: ArcFace loss (additive angular margin)

5:13Step 3: Occlusion augmentation & identification accuracy

8:29Step 4: Training three models

13:58Step 5: Training results

14:31Done!

Ask the class material

My notes

Questions & answers

Your instructor

Introduction

Step 1: Multi-scale fusion & single-scale comparison

Step 2: ArcFace loss (additive angular margin)

Step 3: Occlusion augmentation & identification accuracy

Step 4: Training three models

Step 5: Training results

Done!

Introduction

Step 1: Multi-scale fusion & single-scale comparison

Step 2: ArcFace loss (additive angular margin)

Step 3: Occlusion augmentation & identification accuracy

Step 4: Training three models

Step 5: Training results

Done!