Meeting Recap

Date & Time: 3 April 2025 2PM

Attendees:

• Yoo, Y

• Noah GOSCINIAK

---

1. Executive Summary

Yoo, Young Joon shared expert insights on modern computer vision segmentation techniques. Key recommendations included U-Net variants and Fully Convolutional Network (FCN) deep learning models to address our image segmentation needs.

---

2. Discussion Points

Computer Vision Segmentation Overview

• Importance of precise pixel-level classification for our use cases (e.g., medical imagery, scene understanding).

U-Net Family of Models

• Architecture: Encoder–decoder with skip connections for fine-grained feature recovery.

• Variants:

  • U-Net++ (nested U-Net) for richer multi-scale feature fusion.

  • Attention U-Net to focus on relevant spatial regions.

• Pros & Cons:

  • • Excellent performance with limited data.

  • – Higher memory footprint; may require patch-based training.

Fully Convolutional Networks (FCNs)

• Core Idea: Replace fully connected layers with convolutional layers for end-to-end segmentation.

• Notable Models:

  • FCN-8s for coarse-to-fine upsampling.

  • DeepLab (atrous convolutions + CRF) for sharper boundaries.

• Pros & Cons:

  • • Simpler architecture & faster inference.

  • – Can struggle with small object segmentation without additional post-processing.

Model Selection Criteria

• Data Volume & Annotation Quality

• Compute Resources & Inference Latency

• Target Application Requirements (e.g., boundary precision vs. speed)

---

3. Decision

Evaluate and prototype both a U-Net variant and an FCN-based model on our dataset to compare accuracy, speed, and resource utilization.

---

4. Next Steps & Action Items

Set up data pipeline and preprocess images for segmentation tasks Implement baseline U-Net and run initial training/validation Implement baseline FCN-8s (and DeepLab if resources allow) Compare metrics (IoU, pixel accuracy, inference time) Review results with Yoo, Young Joon and refine model selection