an image segmentation model designed by me using the swin-v2-t transformer classifier as a backbone, a feature pyramid network, lion optimizer, and deep supervised training (using auxiliary layers for loss)
tried using boundary refinement module (sobel kernel) to improve IoU score (0.66 on pascal voc 2012) but i think it may have just added noise:
with boundary-module (purple), without (blue)
training for swin-based fpn segmentation model: experienced irrecoverable loss at 2k steps. Pixel accuracy: 0.75, miou: 0.25
