DLSU Center for Computational Imaging and Visual Innovations

🎉 New Publication Alert! 🎉

We are proud to share that our latest paper has been published in IEEE Access, marking CIVI's first Q1 publication!

🧠 Title: Channeling Fairness: Class Imbalance-Aware Skin Disease Recognition via Fair Channel Enhancement Module
🔗 Read the paper here: https://ieeexplore.ieee.org/abstract/document/11407488

Skin disease classification is particularly challenging due to class imbalance, low inter-class variability, and high intra-class variation in clinical datasets. To address these issues, we proposed Fair Channel Enhancement (FCE), a simple yet powerful module that improves fine-grained feature representation without additional annotations or complex architectures.

🔬 Key Contributions & Results:
• Allocates feature channels proportionally based on class frequency for fairer representation
• Combined with CutMix and label smoothing for improved robustness
• Achieves up to 7.13% accuracy improvement over baseline models
• Improves both low- and high-frequency class accuracy by up to 8.60% and 10.48%, respectively
• Generalizes well to other medical datasets, including ISIC 2018 and Hyper-Kvasir

Our results demonstrate that FCE provides a practical and effective solution for imbalanced medical image classification, which a persistent challenge in real-world AI healthcare applications.

This milestone reflects the dedication of our research team and our continued commitment to advancing fair, robust, and impactful computer vision systems in healthcare.

Onward to more breakthroughs. 🚀

🏅 Outstanding Thesis Award – Silver Medal 🏅

Congratulations to Charles Joseph Hinolan, Mac Andre Javellana, Mari Salvador Lapuz, and Audrea Arjaemi Tabadero for receiving the Outstanding Thesis Award (Silver Medal) for their work titled: "A Lightweight Computer Vision Model for Camouflaged Crop Detection."

This marks the 2nd Outstanding Thesis Award of our lab in the BSCS program and our 3rd award overall — a testament to the consistency and impact of our students' research.

🌱 About the Research
Detecting camouflaged crops remains a major challenge in agricultural computer vision, especially when deploying models on resource-constrained edge devices.

The team conducted a systematic evaluation of lightweight design strategies, such as backbone replacement, pruning, and knowledge distillation across SSD, YOLOv8l, and RT-DETR architectures using various datasets. Their findings highlight key trade-offs between detection accuracy (mAP) and computational efficiency, offering practical insights for real-time agricultural deployment.

This work contributes meaningful guidance toward optimizing object detection systems for real-world agricultural environments where efficiency and accuracy must go hand in hand.

We are incredibly proud of the team for pushing the boundaries of applied computer vision in agriculture. 🚜📊

RESEARCH MODE: ON. ✍️🔥

The DLSU Center for Computational Imaging and Visual Innovations, under the DLSU Advanced Research Institute for Informatics, Computing and Networking, successfully held Focused Research and Manuscript Enhancement (FRAME) Sprint – Day 1 at the John Gokongwei Innovation Center, DLSU Laguna Campus, on January 27, 2026 for BSMSCS students and on February 16, 2026 for THS-ST3 students.

The FRAME Sprint is designed to provide high-focus time dedicated to strengthening research outputs and refining conference manuscripts, with the goal of making each paper submission-ready for reputable conferences and journals. This will be a recurring event throughout the term.

Through guided feedback, structured writing sessions, and collaborative review, our students worked intensively on sharpening their problem statements, clarifying methodologies, strengthening results analysis, and improving overall technical writing quality.

More drafts polished. More papers closer to submission. 🚀📄

🚀 NEW RESEARCH PROJECT 🚀

We are honored to share that our team participated in the DOST-PCIEERD Grand MOA Signing for 2025 Approved Projects held on December 11, 2025 at the Philippine Trade Training Center.

Our collaborator, DLSU GAME Lab, led by Dr. Neil Patrick Del Gallego, officially signed the Memorandum of Agreement for the project: "Game Development Meets Computer Vision: Using Synthetic 3D Scenes for Computer Vision Tasks"

This project, funded by the Department of Science and Technology - Philippine Council for Industry, Energy and Emerging Technology Research and Development (DOST-PCIEERD), explores how photorealistic virtual environments built using tools such as Blender, Unity, and Unreal Engine can generate high-quality synthetic datasets for deep learning in computer vision.

By constructing diverse 3D scenes with varying lighting conditions and object configurations, the project aims to:
• Supplement or substitute real-world image datasets
• Reduce large-scale data collection costs
• Enable scalable training for object recognition and other computer vision tasks

CIVI is proud to serve as a collaborating lab in this exciting initiative that bridges game development and AI-driven computer vision research.

This MOA signing event is attended by Dr. Enrico Paringit (Executive Director, DOST-PCIEERD), Ms. Edna Nacianceno (Chief Science Research Specialist, DOST-PCIEERD), Dr. Raymond Tan (VP for Research and Innovation, DLSU), Dr. Neil Patrick Del Gallego (Head, DLSU Game Lab), and Mr. Arren Matthew Antioquia (Head, DLSU CIVI).

We look forward to contributing to impactful, innovative research that advances the Philippine AI ecosystem. 🇵🇭

🔬 CIVI AT THE OPEN LAB DAY 2026 🔬

Visit our booth at the ground floor of Gokongwei Hall. The booth is open from 12:00 NN to 5:00 PM, February 18, 2026 (Wednesday). We will discuss different research tasks, activities, key research areas, and potential research topics of our center.

You may also view our lab primer here: https://www.canva.com/design/DAHBj99quCU/6TJH4KSSQ9SukLisTf2Zuw/view?utm_content=DAHBj99quCU&utm_campaign=designshare&utm_medium=link2&utm_source=uniquelinks&utlId=h71703f6ba0

[AY 2526 T1 Graduate 🎓]

Congratulations to our outstanding student, Joseph Thomas Año! 🎉

We are incredibly proud to celebrate his graduation with BS (Honors) in Computer Science and MS in Computer Science.

As a valued member of CIVI, Joseph has made impactful contributions through both research and leadership. He has published a conference paper and is on track to publish a journal paper, which demonstrates his commitment to advancing scientific knowledge. He has also gifted our lab with its first-ever Outstanding Thesis Award - Gold Medal at the BSMSCS level, which is an extraordinary milestone for both Joseph and the lab

🏅 Awards & Distinctions:
• Magna Cum Laude
• Outstanding Thesis Award - Gold Medal
• Jose Rizal Honors Award
• Loyalty Award

His accomplishments - both academic and research - are truly inspiring. We are excited to see where his journey leads next.

Congratulations once again, Joseph! 🎓

DLSU Center for Computational Imaging & Visual Innovations (CIVI)

CIVI 2526T1 Technical Workshop Series
📆 November 19, 2025 (W)
⏰ 1:00 PM
📹 Online via Zoom
📱 Register here: https://forms.gle/D8ZorgGyALjU64zw7

Talk 1: Development of Camouflaged Object Detection Networks
Talk 2: Training a YOLO Model Using the NuImages Dataset
Talk 3: Hands-On Pruning of Computer Vision Models in PyTorch
Talk 4: Implementing Curriculum Learning in PyTorch

Talk 1: Development of Camouflaged Object Detection Networks

ABSTRACT
This hands-on workshop offers participants a comprehensive introduction to Camouflaged Object Detection (COD), a field in computer vision that focuses on segmenting hard-to-spot images in environments. Designed for advanced learners, the session walks through the end-to-end workflow of developing and evaluating COD models. Attendees will be guided through the practical steps of setting up and training COD models, followed by evaluation procedures using standard metrics. The session concludes with an analysis of model performance through quantitative metrics, helping participants interpret detection results and refine their models effectively. By the end of the workshop, participants will have built a functional COD pipeline and gained the confidence to apply these techniques to their own projects.

The outline of the workshop is as follows:
1. Introduction to Object Segmentation
2. Dataset Preparation (Preprocessing + Modelling)
3. Environment Preparation
4. Review of Architecture behind Camouflaged Object Detection Models
5. Setting up the Models
6. Training the Models
7. Evaluating the Models
8. Observing Model Detection metrics and results

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Understanding of Neural Network Concepts (e.g., weights, loss function, backpropagation, etc.)
- Understanding of Convolutional Neural Networks (CNN)
- Basic Knowledge of Computer Vision & Image Recognition Tasks
- Understanding of Residual Learning
- Understanding of Object Segmentation

Dependencies
Dependencies needed for the Jupyter notebook:
- Python >= 3.14 (Anaconda Environment Recommended)
- Pytorch (CUDA Version = 3.9
- Pytorch (CUDA Version Recommended)
- Ultralytics YOLO
- OpenCV
- PyYAML
- Matplotlib
- tqdm

ABOUT THE SPEAKER
Matthew Ryan Carandang is an undergraduate student at De La Salle University, pursuing a degree in Computer Science as part of the honors program BSMSCS. Currently, he is conducting his thesis research on single-stage object detection, specifically developing a module to enhance a model's performance for vehicle detection.

Talk 3: Hands-on Pruning of Computer Vision Models in PyTorch

ABSTRACT
Pruning is one of the most practical techniques for reducing the size and computational cost of deep learning models while preserving accuracy. This workshop gives participants hands-on experience with pruning in PyTorch. Using a small CNN trained on a lightweight dataset, attendees will learn how to apply unstructured and structured pruning, retrain the pruned model, and measure performance trade-offs.

The outline of the workshop is as follows:
1. Training a baseline CNN
2. Applying unstructured pruning in PyTorch
3. Applying structured pruning in PyTorch
4. Fine-tuning and retraining the pruned model
5. Evaluating trade-offs

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Familiarity with CNNs
- Basic experience working with PyTorch

Dependencies
Dependencies needed for the Jupyter notebook:
- Python >= 3.8
- PyTorch (CUDA version recommended)
- Torchvision
- Numpy
- Matplotlib

ABOUT THE SPEAKER
Rafael Subo A. Yap is currently a 3rd-year Computer Science student at De La Salle University (DLSU) under the BSMSCS Honors program. He is a tutor with the Peer Tutors Society (PTS) and a member of the Center for Computational Imaging & Visual Innovations (CIVI). His current research focuses on the optimization of deep learning models for counting small objects, with a particular interest in mosquito egg detection as a tool for vector surveillance and public health.

Talk 4: Implementing Curriculum Learning in PyTorch

ABSTRACT
Curriculum learning (CL) is a learning paradigm that trains machine learning models the same way that humans learn, by starting with easier problems and gradually increasing the difficulty. This strategy has shown to improve generalization and convergence rate for a variety of models in computer vision and other fields. This workshop will go over the steps to implement CL using PyTorch. Both pre-defined and self-paced methods will be covered.

The outline of the workshop is as follows:
1. Overview of CL architecture
2. Implementing pre-defined CL
3. Implementing self-paced learning
4. Comparing CL performance with standard methods
5. Discussing results

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Basic understanding of machine learning concepts
- Have an understanding of CNN architectures
- Has experience working with PyTorch

Dependencies
Dependencies needed for the Jupyter notebook:
- Python >= 3.8
- Pytorch (CUDA Version Recommended)
- Torchvision

ABOUT THE SPEAKER
Christian V. Tia is an undergraduate student in the BSMS Computer Science program at De La Salle University (DLSU). His research focuses on automated diabetic retinopathy classification and how it can be optimized in low-resource settings using smartphone-based fundus imaging.

DLSU Center for Computational Imaging & Visual Innovations (CIVI)

CIVI 2526T1 Technical Workshop Series
📆 November 19, 2025 (W)
⏰ 1:00 PM
📹 Online via Zoom
📱 Register here: https://forms.gle/D8ZorgGyALjU64zw7

Talk 1: Development of Camouflaged Object Detection Networks
Talk 2: Training a YOLO Model Using the NuImages Dataset
Talk 3: Hands-On Pruning of Computer Vision Models in PyTorch
Talk 4: Implementing Curriculum Learning in PyTorch

Talk 1: Development of Camouflaged Object Detection Networks

ABSTRACT
This hands-on workshop offers participants a comprehensive introduction to Camouflaged Object Detection (COD), a field in computer vision that focuses on segmenting hard-to-spot images in environments. Designed for advanced learners, the session walks through the end-to-end workflow of developing and evaluating COD models. Attendees will be guided through the practical steps of setting up and training COD models, followed by evaluation procedures using standard metrics. The session concludes with an analysis of model performance through quantitative metrics, helping participants interpret detection results and refine their models effectively. By the end of the workshop, participants will have built a functional COD pipeline and gained the confidence to apply these techniques to their own projects.

The outline of the workshop is as follows:
1. Introduction to Object Segmentation
2. Dataset Preparation (Preprocessing + Modelling)
3. Environment Preparation
4. Review of Architecture behind Camouflaged Object Detection Models
5. Setting up the Models
6. Training the Models
7. Evaluating the Models
8. Observing Model Detection metrics and results

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Understanding of Neural Network Concepts (e.g., weights, loss function, backpropagation, etc.)
- Understanding of Convolutional Neural Networks (CNN)
- Basic Knowledge of Computer Vision & Image Recognition Tasks
- Understanding of Residual Learning
- Understanding of Object Segmentation

Dependencies
Dependencies needed for the Jupyter notebook:
- Python >= 3.14 (Anaconda Environment Recommended)
- Pytorch (CUDA Version = 3.9
- Pytorch (CUDA Version Recommended)
- Ultralytics YOLO
- OpenCV
- PyYAML
- Matplotlib
- tqdm

ABOUT THE SPEAKER
Matthew Ryan Carandang is an undergraduate student at De La Salle University, pursuing a degree in Computer Science as part of the honors program BSMSCS. Currently, he is conducting his thesis research on single-stage object detection, specifically developing a module to enhance a model's performance for vehicle detection.

Talk 3: Hands-on Pruning of Computer Vision Models in PyTorch

ABSTRACT
Pruning is one of the most practical techniques for reducing the size and computational cost of deep learning models while preserving accuracy. This workshop gives participants hands-on experience with pruning in PyTorch. Using a small CNN trained on a lightweight dataset, attendees will learn how to apply unstructured and structured pruning, retrain the pruned model, and measure performance trade-offs.

The outline of the workshop is as follows:
1. Training a baseline CNN
2. Applying unstructured pruning in PyTorch
3. Applying structured pruning in PyTorch
4. Fine-tuning and retraining the pruned model
5. Evaluating trade-offs

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Familiarity with CNNs
- Basic experience working with PyTorch

Dependencies
Dependencies needed for the Jupyter notebook:
- Python >= 3.8
- PyTorch (CUDA version recommended)
- Torchvision
- Numpy
- Matplotlib

ABOUT THE SPEAKER
Rafael Subo A. Yap is currently a 3rd-year Computer Science student at De La Salle University (DLSU) under the BSMSCS Honors program. He is a tutor with the Peer Tutors Society (PTS) and a member of the Center for Computational Imaging & Visual Innovations (CIVI). His current research focuses on the optimization of deep learning models for counting small objects, with a particular interest in mosquito egg detection as a tool for vector surveillance and public health.

Talk 4: Implementing Curriculum Learning in PyTorch

ABSTRACT
Curriculum learning (CL) is a learning paradigm that trains machine learning models the same way that humans learn, by starting with easier problems and gradually increasing the difficulty. This strategy has shown to improve generalization and convergence rate for a variety of models in computer vision and other fields. This workshop will go over the steps to implement CL using PyTorch. Both pre-defined and self-paced methods will be covered.

The outline of the workshop is as follows:
1. Overview of CL architecture
2. Implementing pre-defined CL
3. Implementing self-paced learning
4. Comparing CL performance with standard methods
5. Discussing results

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Basic understanding of machine learning concepts
- Have an understanding of CNN architectures
- Has experience working with PyTorch

Dependencies
Dependencies needed for the Jupyter notebook:
- Python >= 3.8
- Pytorch (CUDA Version Recommended)
- Torchvision

ABOUT THE SPEAKER
Christian V. Tia is an undergraduate student in the BSMS Computer Science program at De La Salle University (DLSU). His research focuses on automated diabetic retinopathy classification and how it can be optimized in low-resource settings using smartphone-based fundus imaging.

DLSU Center for Computational Imaging & Visual Innovations (CIVI)

CIVI 2526T1 Academic Lecture Series
📆 November 12, 2025 (W)
⏰ 1:00 PM
📹 Online via Zoom
📱 Register here: https://forms.gle/W9LxYpgpKrSn39HK9

Talk 1: An Introduction to Camouflaged Object Detection
Talk 2: Introduction to Single-Stage Object Detection
Talk 3: Pruning in Computer Vision: Making Object Detection Lighter
Talk 4: Curriculum Learning for Image Classification

Talk 1: An Introduction to Camouflaged Object Detection
ABSTRACT
This lecture will cover Camouflaged Object Detection (COD), a field of Computer Vision that focuses on segmenting camouflaged objects hidden in images. Designed for advanced learners, this lecture will cover the history and prominent trends in COD, as well as the notable networks and metrics used in the field.

The outline of the lecture is as follows:
1. Definition of COD
2. History and Creation of COD
3. COD Network Architectures
4. COD Performance Metrics
5. Current Applications for COD

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Understanding of Neural Network Concepts (e.g., weights, loss function, backpropagation, etc.)
- Understanding of Convolutional Neural Networks (CNN)
- Basic Knowledge of Computer Vision
- Understanding of Residual Learning
- Understanding of Object Segmentation

ABOUT THE SPEAKER
Aaron Gabrielle C. Dichoso is currently a 4th-year Computer Science student at De La Salle University (DLSU) under the BSMSCS Honors program. As a member of CIVI, his laboratory experiments focus on Camouflaged Crop Detection (CCD), with an interest in pursuing the improvement of CCD Network Architecture for his thesis research in the pursuit of accelerating agricultural infrastructure in the country.

Talk 2: Introduction to Single-Stage Object Detection
ABSTRACT
Single-stage object detectors have become a cornerstone in modern computer vision due to their balance of speed and accuracy. Unlike traditional two-stage detectors (e.g., Faster R-CNN) that separate region proposal and classification, single-stage detectors perform object localization and classification in one pass, making them highly efficient for real-time applications. This lecture provides a comprehensive introduction to single-stage object detection, covering the fundamental principles, popular architectures such as SSD, RetinaNet, and YOLO, as well as recent improvements in anchor-free methods. Attendees will gain insights into the design trade-offs between speed and accuracy, evaluation metrics, and practical use cases in industry and research.

The outline of the lecture is as follows:
1. Introduction to Object Detection
2. Single-Stage Object Detection Architectures
3. Two-stage vs single-stage approaches
4. Advantages and Limitations of Single-Stage Detectors
5. Applications of Single-Stage Object Detectors

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Basic understanding of machine learning concepts
- Familiarity with convolutional neural networks (CNNs)
- Basic knowledge of object detection terminology (bounding boxes, IoU, precision/recall)

ABOUT THE SPEAKER
Matthew Ryan Carandang is an undergraduate student at De La Salle University, pursuing a degree in Computer Science as part of the honors program BSMSCS. Currently, he is conducting his thesis research on single-stage object detection, specifically developing a module to enhance a model's performance for vehicle detection.

Talk 3: Pruning in Computer Vision: Making Object Detection Lighter
ABSTRACT
As deep learning models grow increasingly complex, their deployment for real-world tasks faces challenges in computation and memory. Model optimization techniques like pruning offer practical solutions to reduce model size and improve efficiency while preserving accuracy. This lecture explores the theory of pruning, distinguishing structured and unstructured methods. While applicable to a wide range of models, this lecture highlights case studies involving YOLO object detection models, which are often optimized for mobile and edge deployment.

The outline of the lecture is as follows:
1. Motivation for optimizing object detection models
2. Overview of model compression methods
3. Pruning theory and approaches
4. Trade-offs in pruning: accuracy vs. speed vs. size

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Basic understanding of deep learning models
- General knowledge of object detection

ABOUT THE SPEAKER
Rafael Subo A. Yap is currently a 3rd-year Computer Science student at De La Salle University (DLSU) under the BSMSCS Honors program. He is a tutor with the Peer Tutors Society (PTS) and a member of the Center for Computational Imaging & Visual Innovations (CIVI). His current research focuses on the optimization of deep learning models for counting small objects, with a particular interest in mosquito egg detection as a tool for vector surveillance and public health.

Talk 4: Curriculum Learning for Image Classification
ABSTRACT
Curriculum learning (CL) is a learning paradigm that trains machine learning models the same way that humans learn, by starting with easier problems and gradually increasing the difficulty. This strategy has shown to improve generalization and convergence rate for a variety of models in computer vision and other fields. This lecture will cover the core concepts behind CL, including a review of its history, variations, and applications.

The outline of the lecture is as follows:
1. Introduction and motivations
2. Definitions of CL
3. Theoretical basis
4. General framework
5. Potential Applications
6. Conclusion

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Basic understanding of machine learning concepts
- Have an understanding of CNN architectures

ABOUT THE SPEAKER
Christian V. Tia is an undergraduate student in the BSMS Computer Science program at De La Salle University (DLSU). His research focuses on automated diabetic retinopathy classification and how it can be optimized in low-resource settings using smartphone-based fundus imaging.

DLSU Center for Computational Imaging & Visual Innovations (CIVI)

CIVI 2526T1 Academic Lecture Series
📆 November 12, 2025 (W)
⏰ 1:00 PM
📹 Online via Zoom
📱 Register here: https://forms.gle/W9LxYpgpKrSn39HK9

Talk 1: An Introduction to Camouflaged Object Detection
Talk 2: Introduction to Single-Stage Object Detection
Talk 3: Pruning in Computer Vision: Making Object Detection Lighter
Talk 4: Curriculum Learning for Image Classification

Talk 1: An Introduction to Camouflaged Object Detection
ABSTRACT
This lecture will cover Camouflaged Object Detection (COD), a field of Computer Vision that focuses on segmenting camouflaged objects hidden in images. Designed for advanced learners, this lecture will cover the history and prominent trends in COD, as well as the notable networks and metrics used in the field.

The outline of the lecture is as follows:
1. Definition of COD
2. History and Creation of COD
3. COD Network Architectures
4. COD Performance Metrics
5. Current Applications for COD

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Understanding of Neural Network Concepts (e.g., weights, loss function, backpropagation, etc.)
- Understanding of Convolutional Neural Networks (CNN)
- Basic Knowledge of Computer Vision
- Understanding of Residual Learning
- Understanding of Object Segmentation

ABOUT THE SPEAKER
Aaron Gabrielle C. Dichoso is currently a 4th-year Computer Science student at De La Salle University (DLSU) under the BSMSCS Honors program. As a member of CIVI, his laboratory experiments focus on Camouflaged Crop Detection (CCD), with an interest in pursuing the improvement of CCD Network Architecture for his thesis research in the pursuit of accelerating agricultural infrastructure in the country.

Talk 2: Introduction to Single-Stage Object Detection
ABSTRACT
Single-stage object detectors have become a cornerstone in modern computer vision due to their balance of speed and accuracy. Unlike traditional two-stage detectors (e.g., Faster R-CNN) that separate region proposal and classification, single-stage detectors perform object localization and classification in one pass, making them highly efficient for real-time applications. This lecture provides a comprehensive introduction to single-stage object detection, covering the fundamental principles, popular architectures such as SSD, RetinaNet, and YOLO, as well as recent improvements in anchor-free methods. Attendees will gain insights into the design trade-offs between speed and accuracy, evaluation metrics, and practical use cases in industry and research.

The outline of the lecture is as follows:
1. Introduction to Object Detection
2. Single-Stage Object Detection Architectures
3. Two-stage vs single-stage approaches
4. Advantages and Limitations of Single-Stage Detectors
5. Applications of Single-Stage Object Detectors

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Basic understanding of machine learning concepts
- Familiarity with convolutional neural networks (CNNs)
- Basic knowledge of object detection terminology (bounding boxes, IoU, precision/recall)

ABOUT THE SPEAKER
Matthew Ryan Carandang is an undergraduate student at De La Salle University, pursuing a degree in Computer Science as part of the honors program BSMSCS. Currently, he is conducting his thesis research on single-stage object detection, specifically developing a module to enhance a model's performance for vehicle detection.

Talk 3: Pruning in Computer Vision: Making Object Detection Lighter
ABSTRACT
As deep learning models grow increasingly complex, their deployment for real-world tasks faces challenges in computation and memory. Model optimization techniques like pruning offer practical solutions to reduce model size and improve efficiency while preserving accuracy. This lecture explores the theory of pruning, distinguishing structured and unstructured methods. While applicable to a wide range of models, this lecture highlights case studies involving YOLO object detection models, which are often optimized for mobile and edge deployment.

The outline of the lecture is as follows:
1. Motivation for optimizing object detection models
2. Overview of model compression methods
3. Pruning theory and approaches
4. Trade-offs in pruning: accuracy vs. speed vs. size

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Basic understanding of deep learning models
- General knowledge of object detection

ABOUT THE SPEAKER
Rafael Subo A. Yap is currently a 3rd-year Computer Science student at De La Salle University (DLSU) under the BSMSCS Honors program. He is a tutor with the Peer Tutors Society (PTS) and a member of the Center for Computational Imaging & Visual Innovations (CIVI). His current research focuses on the optimization of deep learning models for counting small objects, with a particular interest in mosquito egg detection as a tool for vector surveillance and public health.

Talk 4: Curriculum Learning for Image Classification
ABSTRACT
Curriculum learning (CL) is a learning paradigm that trains machine learning models the same way that humans learn, by starting with easier problems and gradually increasing the difficulty. This strategy has shown to improve generalization and convergence rate for a variety of models in computer vision and other fields. This lecture will cover the core concepts behind CL, including a review of its history, variations, and applications.

The outline of the lecture is as follows:
1. Introduction and motivations
2. Definitions of CL
3. Theoretical basis
4. General framework
5. Potential Applications
6. Conclusion

Prerequisite knowledge
Attendees should be knowledgeable about these topics:
- Basic understanding of machine learning concepts
- Have an understanding of CNN architectures

ABOUT THE SPEAKER
Christian V. Tia is an undergraduate student in the BSMS Computer Science program at De La Salle University (DLSU). His research focuses on automated diabetic retinopathy classification and how it can be optimized in low-resource settings using smartphone-based fundus imaging.