Xinghao Chen

Xinghao Chen

M.S. Student

CostalSeg Coastal Erosion Platform

Overview

CostalSeg is an applied project of University of Washington that builds an end-to-end workflow for community-driven coastal erosion monitoring. The toolkit ingests user photos, filters low-quality submissions, aligns multi-angle imagery, and produces seven-class segmentations powered by a DeepLabV3+ model with an EfficientNet-B6 backbone. The production checkpoints reach a 0.93 IoU on the program’s labeled datasets and integrate directly with the Applied Physics Laboratory’s MyCoast submission flow.

My Contributions

  • Led model training and evaluation for the Metal Marcy and Silhouette Jaenette beaches, including data preparation scripts and packaged pretrained checkpoints.
  • Implemented the Gradio-based browser app (app.py) that wraps inference, overlay rendering, and batch processing for field researchers.
  • Authored documentation for model retraining, outlier detection, and deployment to Hugging Face Spaces.

Highlights

  • Multi-class segmentation pipeline - DeepLabV3Plus + EfficientNet-B6 models exported through segmentation_models_pytorch; utilities for both scratch training and loading shared checkpoints.
  • Interactive analyst UI - One-click Gradio experience with layered visualization, downloadable masks, and batch folder processing.
  • Outlier detection guardrail - ViT-H-14 feature embeddings compare incoming photos against curated references to keep the training corpus clean.
  • Perspective correction - SuperGlue keypoint matching plus homography warping keeps shoreline contours consistent across community uploads.
  • Performance tooling - model_performance_test.py records inference throughput, GPU utilization, and FLOP efficiency across batch sizes and image resolutions.
  • Hosted demos - Three Hugging Face Spaces (MetalMarcy, SilhouetteJaenette, CostalSegment) expose the workflow to project partners without local setup.

Technical Stack

Python, PyTorch, segmentation_models_pytorch, Gradio, Hugging Face transformers, OpenCV, NumPy

Learn More