Image Segmentation Model for Photos

I have a collection of everyday photographs and I need a reliable computer-vision solution that can segment each image into precise pixel-level masks. The end goal is a fully trained image-segmentation model that I can drop into my existing pipeline and run locally or in the cloud.

Here is what I expect:
• A well-documented data-preprocessing script that loads general photographs, applies standard augmentations, and prepares train/val/test splits.
• Model-training code in PyTorch or TensorFlow, with clear comments and a configurable YAML/JSON file for hyper-parameters.
• A set of pretrained weights that achieve competitive Intersection-over-Union (IoU) scores on the validation set.
• An inference script (CLI or notebook) that accepts a single image or a folder of images and returns segmentation masks.
• Export of the final network to ONNX or TensorFlow SavedModel so it can be served easily.
• README explaining environment setup (CUDA versions, required Python packages), training commands, and how to reproduce the reported metrics.

If you prefer alternative architectures—U-Net, DeepLab, Mask R-CNN, or any transformer-based option—feel free to go with what you know best as long as it delivers clean, accurate masks. Please keep the code neatly organized, adhere to best practices, and make sure the solution runs on both GPU and CPU.

When the above items are complete and tested, the project is done.