Train a segmentation model to detect rooflines in house photos

Budget: $300 fixed price I need a trained computer vision model that takes a photo of a house and outputs the roofline as a mask I can convert to a polyline. This feeds a drawing tool for holiday lighting mockups. Scope: 1. Label a dataset of house photos. I will supply photos and a written labeling spec. SAM assisted labeling is fine and encouraged. 2. Train a lightweight segmentation model on that dataset. SegFormer-B0, a compact UNet, or similar. Your call, tell me why. 3. Iterate until it performs on a holdout set of 30 photos you will not see until delivery. Deliverables: • Trained model weights • Training script, so I can retrain later without you • Inference code as a Python function: image in, mask out • The labeled dataset Not in scope: ONNX export, browser inference, deployment, geometry, UI, pricing logic. I handle all of that. How I score it: I have 30 photos with the correct roofline marked by me. You never see them. I run your model on them and compare. That is the acceptance test, nothing else. To apply, answer these three: 1. What segmentation architecture would you use here and why that one? 2. How many labeled images do you think this needs? 3. Link one project where you trained a segmentation model yourself, not fine tuned an API. Applications that do not answer all three will not be read. If your plan is to call a hosted vision API per image, this is not the job.