Image segmentation: landslides and VOC
Segmentation experiments and architecture analysis
Hands-on segmentation work covering landslide recognition and VOC2012 experiments with FCN and U-Net models.
ROLE
Model implementation and experimentation
APPROACH
Remote sensing · U-Net · Landslides
Research question
My segmentation work includes landslide-recognition experiments and a separate VOC2012 workflow. I connect architectural reading with model training, validation metrics, and prediction visualization.
My contribution
- Ran landslide-recognition experiments and studied multi-scale fusion and attention in segmentation architectures.
- Built and ran a VOC2012 workflow with FCN/U-Net-family implementations, training and validation logs, checkpoints, and prediction visualization.
Technical approach
- Study multi-scale fusion, residual attention, nested skip connections, and reconstruction in landslide segmentation.
- Use VOC2012 image/mask splits for the separate general-image segmentation experiments.
- Track loss, mIoU, pixel accuracy, frequency-weighted IoU and mean accuracy, and inspect predicted masks.
Results & outcomes
- Completed practical segmentation experiments alongside the architectural study.
- The VOC workflow includes saved model checkpoints, training and validation logs, loss comparisons, and image-level prediction outputs.
