Jekyll2026-03-05T22:05:44+09:00https://yakhyo.github.io/feed.xmlHomeExploring Life Through the Lens of a Computer Scientist: AI, Tech, and Beyond.Yakhyokhuja ValikhujaevUniFace: All-in-One Face Analysis Toolkit for Production2025-11-11T12:00:00+09:002025-11-11T12:00:00+09:00https://yakhyo.github.io/blog/2025/11/uniface-all-in-one-face-analysisUniFace is a lightweight, production-ready face analysis library built on ONNX Runtime. It provides a unified API for face detection, recognition, landmark detection, attribute analysis, face parsing, gaze estimation, anti-spoofing, and privacy features.
Documentation Moved
The comprehensive documentation for UniFace has been moved to a dedicated documentation site with tutorials, API references, and guides.
]]>Yakhyokhuja ValikhujaevFace Parsing using BiSeNet for Real-time Semantic Segmentation2024-11-29T12:00:00+09:002024-11-29T12:00:00+09:00https://yakhyo.github.io/blog/2024/11/face-parsing-bisenetBiSeNet (Bilateral Segmentation Network) is a state-of-the-art model for real-time semantic segmentation, initially proposed in the paper Bilateral Segmentation Network for Real-time Semantic Segmentation. The architecture addresses the fundamental challenge in semantic segmentation: achieving high accuracy while maintaining real-time performance.
Architecture Overview
BiSeNet combines two complementary paths to balance spatial detail and semantic context:
Spatial Path: Preserves high-resolution spatial information through a shallow network with wide channels, capturing fine-grained details essential for precise segmentation boundaries.
Context Path: Employs a lightweight backbone to aggregate rich contextual information with a large receptive field, enabling accurate semantic understanding.
The fusion of these paths through a Feature Fusion Module ensures high segmentation accuracy with low computational cost, making it ideal for applications requiring real-time performance on resource-constrained devices.
Key Features
Accurate Facial Parsing: Segments detailed facial features including eyes, nose, mouth, and hair for precise analysis
ONNX Support: Seamless conversion from PyTorch to ONNX format for cross-platform deployment
Flexible Backbones: Support for ResNet18 and ResNet34, allowing trade-offs between speed and accuracy
Production-Ready: Optimized for real-time applications in AR/VR, digital makeup, and facial analysis systems
Performance Comparison
ResNet34 Backbone
ResNet18 Backbone
Getting Started
Clone the repository and install dependencies:
git clone https://github.com/yakhyo/face-parsing.git
cd face-parsing
pip install-r requirements.txt
Digital Makeup Applications: Precise facial feature segmentation for virtual makeup try-on
Face Swapping: Accurate face region extraction for deepfake and face replacement systems
Facial Analysis: Detailed feature extraction for emotion recognition and facial attribute analysis
AR/VR Applications: Real-time face parsing for augmented reality filters and effects
Visit the GitHub repository for detailed documentation, training scripts, and inference examples.
]]>Yakhyokhuja ValikhujaevTiny-Face: Ultra-lightweight Face Detection for Mobile and Edge Devices2024-11-09T12:00:00+09:002024-11-09T12:00:00+09:00https://yakhyo.github.io/blog/2024/11/efficient-tiny-face-detectorTiny-Face is an ultra-lightweight face detection model specifically designed for deployment on mobile and edge devices where computational resources are limited. Unlike conventional face detection models that prioritize accuracy at the cost of model size and inference speed, Tiny-Face achieves an optimal balance between detection performance and computational efficiency.
Building upon the core concepts of RetinaFace, Tiny-Face introduces several key optimizations that make it practical for real-world deployment on mobile phones, embedded systems, and IoT devices. The model is streamlined to use minimal memory and processing power while maintaining high precision in face detection across various challenging conditions.
Ultra-lightweight Architecture: Model sizes ranging from 1.4MB to 1.8MB, ideal for mobile deployment
Multiple Configurations: SlimFace, RFB, and MobileNet variants optimized for different resource constraints
Real-time Performance: Achieves real-time inference on mobile CPUs without GPU acceleration
Pretrained Models: Ready-to-use weights trained on WiderFace dataset for immediate deployment
Performance on WiderFace Dataset
Multi-scale Image Size
Models
Pretrained on ImageNet
Easy
Medium
Hard
#Params(M)
Size(MB)
SlimFace
False
79.50%
79.40%
68.36%
0.343
1.4
RFB
False
80.49%
81.51%
75.73%
0.359
1.5
RetinaFace
True
87.69%
86.39%
80.21%
0.426
1.8
Original Image Size
Models
Pretrained on ImageNet
Easy
Medium
Hard
#Params(M)
SlimFace
False
87.10%
84.36%
67.38%
0.343
RFB
False
87.09%
84.61%
69.22%
0.359
RetinaFace
True
90.26%
87.48%
72.85%
0.426
Technical Implementation
The model architecture incorporates several optimization techniques:
Depthwise Separable Convolutions: Reduces computational cost while maintaining representational power
Feature Pyramid Network: Multi-scale feature extraction for detecting faces of various sizes
Efficient Anchor Design: Optimized anchor boxes specifically tuned for face detection tasks
Quantization-Friendly: Architecture designed to maintain accuracy after INT8 quantization
Use Cases
Tiny-Face is particularly well-suited for:
Mobile Applications: Face detection in camera apps, social media filters, and photo editing tools
Edge Computing: Real-time face detection on IoT devices and smart cameras
Embedded Systems: Integration into resource-constrained hardware for access control and monitoring
Offline Applications: Face detection without requiring cloud connectivity or GPU acceleration
Explore the GitHub repository for detailed setup instructions, training scripts, and deployment examples for various platforms including Android, iOS, and embedded Linux systems.
]]>Yakhyokhuja ValikhujaevRetinaFace: Single-stage Dense Face Localisation in the Wild2024-10-28T12:00:00+09:002024-10-28T12:00:00+09:00https://yakhyo.github.io/blog/2024/10/high-performance-retinaface-detectorRetinaFace is a robust single-stage face detection framework designed for dense face localisation in unconstrained environments. This implementation provides a production-ready solution with multiple backbone options, enabling flexible deployment across different hardware constraints and accuracy requirements.
The model excels at detecting faces across extreme variations in scale, pose, and occlusion, making it particularly effective for real-world applications where faces may appear at any size or orientation within the image.
UniFace Library: For easier integration, check out UniFace, a lightweight Python library built on models from this repository. UniFace provides a simple API for face detection, alignment, and landmark extraction.
The implementation supports various deployment scenarios:
Python Inference: Direct PyTorch inference for development and testing
ONNX Runtime: Cross-platform deployment with optimized inference
Mobile Deployment: Lightweight MobileNet backbones for on-device inference
Server Deployment: High-accuracy ResNet backbones for cloud-based services
For detailed documentation on training custom models, fine-tuning on specific datasets, and deployment guides, visit the GitHub repository.
]]>Yakhyokhuja ValikhujaevUnderstanding the Geometric Perspective of Vectors in Machine Learning2024-09-19T12:00:00+09:002024-09-19T12:00:00+09:00https://yakhyo.github.io/blog/2024/09/math-for-machine-learningWhile I have a strong mathematical background in calculus and linear algebra, including the ability to perform matrix operations and transformations by hand, I recently discovered a crucial gap in my understanding: the geometric interpretation of vectors and their operations.
This realization came as I began exploring geometric explanations of linear algebra concepts. What were previously abstract lists of numbers or array indices suddenly transformed into intuitive geometric objects—arrows pointing to specific locations in space, representing directions and magnitudes in a visually comprehensible way.
Why Geometric Intuition Matters
Understanding vectors geometrically fundamentally changes how we approach machine learning problems:
Feature Spaces: Each data point becomes a position in high-dimensional space, making concepts like similarity and distance intuitive
Transformations: Matrix operations become geometric transformations—rotations, scalings, and projections
Optimization: Gradient descent transforms from abstract calculus into following the steepest downhill path in a landscape
Embeddings: Word embeddings and latent representations become geometric relationships where similar concepts cluster together
This geometric perspective provides deeper insight into why certain machine learning algorithms work and how to debug them when they don’t.
Recommended Learning Resources
For anyone working in machine learning who hasn’t yet explored the geometric foundations, I highly recommend these courses:
Linear Algebra
Essence of Linear Algebra by 3Blue1Brown - Course Link
An exceptional visual introduction to linear algebra concepts with stunning animations that build geometric intuition
Mathematics for Machine Learning: Linear Algebra - Coursera
Specifically designed for ML practitioners, covering the essential linear algebra needed for understanding modern ML algorithms
Calculus and Optimization
Mathematics for Machine Learning: Multivariate Calculus - Coursera
Focuses on the calculus concepts most relevant to machine learning, particularly gradient-based optimization
Dimensionality Reduction
Mathematics for Machine Learning: PCA - Coursera
Explores Principal Component Analysis from both mathematical and geometric perspectives
Practical Impact
This geometric understanding has practical implications for machine learning work:
Better Model Design: Understanding how transformations affect data helps in designing better architectures
Debugging: Geometric intuition makes it easier to identify why models fail on certain data
Feature Engineering: Knowing how features interact geometrically guides better feature design
Interpretability: Geometric perspective aids in explaining model decisions to non-technical stakeholders
Investing time in building this geometric intuition is one of the most valuable things you can do as a machine learning practitioner. It transforms machine learning from a collection of algorithms and formulas into an intuitive, visual discipline where you can reason about what models are doing and why.
]]>Yakhyokhuja ValikhujaevReal-Time Gaze Estimation Using Lightweight Deep Learning Models2024-09-18T12:00:00+09:002024-09-18T12:00:00+09:00https://yakhyo.github.io/blog/2024/09/gaze-estimationThis project focuses on predicting gaze direction using lightweight deep learning models optimized for real-time performance on mobile devices. The implementation combines classification and regression techniques to create an efficient and accurate solution suitable for deployment on resource-constrained hardware.
Gaze estimation technology enables a wide range of applications across multiple domains:
Mobile User Experience: Hands-free navigation and attention-aware interfaces
Virtual and Augmented Reality: Natural interaction through eye tracking in VR/AR systems
Accessibility: Assistive technologies for users with limited mobility
Automotive Safety: Driver attention monitoring and drowsiness detection
Human-Computer Interaction: Intuitive control mechanisms for various devices
Market Research: Understanding user attention patterns and visual behavior
Model Architecture and Design
The project implements multiple lightweight architectures, each optimized for different deployment scenarios:
ResNet Variants
Employs residual learning techniques to enable deeper networks without degradation. The residual connections allow gradients to flow more effectively during training, resulting in better accuracy without significant computational overhead.
MobileNet v2
Specifically designed for mobile deployment, MobileNet v2 introduces inverted residual structures and linear bottlenecks. This architecture achieves an optimal balance between model size, inference speed, and accuracy, making it ideal for on-device gaze estimation.
MobileOne (s0-s4)
The MobileOne family represents the state-of-the-art in mobile-optimized architectures. With variants ranging from s0 to s4, it offers flexibility in trading off between speed and accuracy. The architecture is specifically optimized for mobile CPUs, achieving impressive real-time performance without GPU acceleration.
Face Detection Integration
The system integrates SCRFD (Sample and Computation Redistribution for Efficient Face Detection) for robust face localization. SCRFD provides:
Fast inference suitable for real-time applications
High accuracy across various face scales and poses
Efficient resource utilization for mobile deployment
Reliable performance in challenging lighting conditions
Technical Implementation
The gaze estimation pipeline consists of several stages:
Face Detection: SCRFD localizes faces in the input frame
Face Alignment: Detected faces are normalized to a standard pose
Eye Region Extraction: Precise localization of eye regions for gaze prediction
Gaze Prediction: Deep learning model estimates gaze direction as pitch and yaw angles
Temporal Smoothing: Optional filtering to reduce jitter in video streams
Performance Characteristics
The implementation achieves:
Real-time inference (30+ FPS) on modern mobile devices
Low latency suitable for interactive applications
Minimal battery impact through efficient computation
Robust performance across different lighting conditions and head poses
The complete implementation, including training scripts, pre-trained models, and deployment examples, is available on GitHub.
]]>Yakhyokhuja ValikhujaevReal-Time Head Pose Estimation with Efficient Deep Learning Backbones2024-09-17T12:00:00+09:002024-09-17T12:00:00+09:00https://yakhyo.github.io/blog/2024/09/head-pose-estimationThis project delivers accurate real-time head pose estimation through optimized deep learning architectures. The implementation focuses on achieving high performance across various deployment scenarios, from mobile devices to desktop applications, while maintaining robust accuracy in challenging conditions.
Head pose estimation plays a critical role in numerous applications:
Augmented and Virtual Reality: Natural user interaction through head tracking
Attention Monitoring: Understanding user focus in educational and workplace settings
Driver Safety Systems: Detecting driver distraction and drowsiness in automotive applications
Human-Computer Interaction: Enabling intuitive control mechanisms
Video Conferencing: Automatic camera adjustment and gaze correction
Accessibility Technologies: Assistive systems for users with limited mobility
Technical Architecture
The system integrates multiple components to achieve robust performance:
Backbone Networks
ResNet Variants
ResNet architectures provide the foundation for high-accuracy head pose estimation. The residual connections enable training of deeper networks, capturing complex patterns in head orientation across various poses and lighting conditions.
MobileNet v2 and v3
MobileNet architectures are specifically optimized for mobile deployment:
