Evgeny Smirnov
A5035017930- Face recognition and analysis
- Face and Expression Recognition
- Biometric Identification and Security
- COVID-19 diagnosis using AI
- Domain Adaptation and Few-Shot Learning
- Advanced Image and Video Retrieval Techniques
- Machine Learning and Data Classification
- Video Surveillance and Tracking Methods
- Speech Recognition and Synthesis
- Music and Audio Processing
- Speech and Audio Processing
- Video Analysis and Summarization
St Petersburg University
2014
Large and Deep Convolutional Neural Networks achieve good results in image classification tasks, but they need methods to prevent overfitting. In this paper we compare performance of different regularization techniques on ImageNet Scale Visual Recognition Challenge 2013. We show empirically that Dropout works better than DropConnect dataset.
In this paper we present Doppelganger mining - a method to learn better face representations. The main idea of is maintain list with the most similar identities for each identity in training set. This used generate mini-batches by sampling pairs similar-looking ("doppelgangers") together. It especially useful methods, based on exemplar-based supervision. Usually hard example comes price necessity use large or substantial extra computation and memory cost, particularly datasets numbers...
Hard example mining is an important part of the deep embedding learning. Most methods perform it at mini-batch level. However, in large-scale settings there only a small chance that proper examples will appear same and be coupled into hard pairs or triplets. Doppelganger was previously proposed to increase this by means class-wise similarity. This method ensures similar classes are sampled together. One drawbacks operates class level, while also might way select appropriate within more...
Mini-batch construction strategy is an important part of the deep representation learning. Different strategies have their advantages and limitations. Usually only one them selected to create mini-batches for training. However, in many cases combination can be more efficient than using them. In this paper, we propose Composite Mini-Batches - a technique combine several mini-batch sampling training process. The main idea compose from parts, use different each part. With kind construction,...
Face representation learning using datasets with a massive number of identities requires appropriate training methods. Softmax-based approach, currently the state-of-the-art in face recognition, its usual "full softmax" form is not suitable for millions persons. Several methods, based on "sampled were proposed to remove this limitation. These however, have set disadvantages. One them problem "prototype obsolescence": classifier weights (prototypes) rarely sampled classes receive too scarce...
Synthetic data is gaining increasing relevance for training machine learning models. This mainly motivated due to several factors such as the lack of real and intra-class variability, time errors produced in manual labeling, some cases privacy concerns, among others. paper presents an overview 2nd edition Face Recognition Challenge Era Data (FRCSyn) organized at CVPR 2024. FRCSyn aims investigate use synthetic face recognition address current technological limitations, including demographic...
Synthetic data is gaining increasing popularity for face recognition technologies, mainly due to the privacy concerns and challenges associated with obtaining real data, including diverse scenarios, quality, demographic groups, among others. It also offers some advantages over such as large amount of that can be generated or ability customize it adapt specific problem-solving needs. To effectively use models should specifically designed exploit synthetic its fullest potential. In order...
In this paper, the robust method for reduction of 'false positive' in<br/>multilevel face detection system is proposed. The based on<br/>convolutional neural network and trained on results from cascades of<br/>boosted classifiers. Moreover, we evaluate effect modern<br/>approaches like Dropout DropConnect learning.
Prototype Memory is a powerful model for face representation learning. It enables the training of recognition models using datasets any size, with on-the-fly generation prototypes (classifier weights) and efficient ways their utilization. demonstrated strong results in many benchmarks. However, algorithm prototype generation, used it, prone to problems imperfectly calculated case low-quality or poorly recognizable faces images, selected creation. All images same person, presented mini-batch,...