WK.

Semantic Segmentation (HOI)

A robust end-to-end Semantic Segmentation pipeline built with PyTorch, focusing on Human Object Interaction. The system utilizes an LR-ASPP MobileNetV3 architecture optimized for consumer GPUs. It features a custom OOM-safe training loop that automatically reduces batch size and image dimensions to prevent memory crashes, along with support for Mixed Precision (AMP) and Gradient Accumulation.

Semantic Segmentation (HOI) - 1
1 / 5

My Role

Computer Vision Engineer — Architecture, Pipeline Development

Tech Stack

PythonPyTorchSemantic SegmentationRoboflowMobileNetV3Computer Vision

Source Code

github.com/wongsakron

Key Highlights

  • LR-ASPP MobileNetV3 architecture optimized for low-memory GPUs
  • OOM-Safe Training Loop with automatic batch and image size adjustment
  • Support for Mixed Precision (AMP) and Gradient Accumulation
  • Integration with Roboflow API for streamlined dataset management
  • Comprehensive custom evaluation metrics (mIoU, Pixel Accuracy, per-class F1-Score)