UperNet Decoder

Purpose: decode UperNet outputs into segmentation maps for validation and visualization. User Intent: ML engineers and researchers who work with semantic segmentation models and need interpretable results for debugging, benchmarking, or feature analysis. Core logic: applies the decoder head’s geometry to map encoded tokens to per-pixel labels, supports multiple backbones and pre/post-processing. Features: accepts encoded maps in base64 or numpy arrays, supports batch processing, configurable post-processing (softmax, thresholding), multiple backbones alignment, and export options (arrays, images). Outputs: per-pixel label maps, optional confidence maps, and optional overlay data with encoding options. Algorithms: relies on standard UperNet decoding steps, combining multi-scale feature maps using the specified fusion strategy, then applying softmax and argmax to produce final segmentation. Edge considerations: validates input shapes, aligns spatial resolutions, handles missing channels, and returns informative error codes when input is incompatible. Localization: adheres to common CV conventions (NCHW/NHWC), supports common data formats and ONNX/torchscript exports, suitable for global research workflows. Assumptions: input is a valid encoded feature map or logits; the tool is used server-side or local; excludes training features or full model inference. Exclusions: training, backprop, dataset creation, UI styling, and any model training capabilities.