Silver Labels Are Not Ground Truth: A Four-Head Fashion Classification Project

In fashion image classification, attribute recognition requires joint estimation across multiple semantic domains. This project structures prediction across four discrete heads: material (6 classes: cotton, denim, knitwear, leather, silk, synthetic), color (12 classes), pattern (6 classes: solid, striped, plaid, floral, geometric, animal print), and texture (5 classes: smooth, rough, fluffy, structured, sheer).

The training and evaluation corpus comprises a curated 13,355-image subset derived from TextileNet. Rather than framing the task as independent single-label problems, a multi-head architecture predicts all four attributes jointly from a shared representation.

Human annotation across thousands of images for multiple fine-grained categories is resource-intensive. To bootstrap training data, the project implemented a two-part weak supervision pipeline: rule-based text mapping for material attributes from product descriptions, and vision-language model (VLM) zero-shot prompting for color, pattern, and texture.

These generated annotations are explicitly treated as silver labels, not verified gold standards. While automated prompting provides rapid dataset coverage, it introduces systematic noise: semantic drift on multi-color garments, ambiguity between tactile texture and geometric pattern, and edge-case hallucination. Low-confidence annotations and ambiguous samples were pruned via thresholded abstention rules before training.

The primary model uses a pre-trained DINOv2 ViT-B/14 backbone with four linear classification heads trained simultaneously. To evaluate feature utility, the project benchmarks DINOv2 against a standard ResNet-50 baseline trained under the identical reported data split, image augmentations, optimizer, cosine learning rate schedule, and class-weight balancing.

Under the documented silver-label test setup, the repository reports a mean macro-F1 of 0.765 for DINOv2 ViT-B/14 versus 0.649 for ResNet-50. Per-head results show consistent advantages for the self-supervised vision transformer: Color (0.827 vs 0.738), Texture (0.753 vs 0.630), Material (0.744 vs 0.612), and Pattern (0.737 vs 0.616).

The project reports a separate stratified human audit of 210 images to inspect the silver-label annotations. This audit is distinct from the model test evaluation: the reported macro-F1 scores measure agreement with the silver test labels, while the audit examines a sampled set of labels against human review.

The public summary does not provide enough detail here to reconstruct the audit sampling and adjudication procedure or to estimate population-wide label error. The two evaluations should therefore remain separate, and the audit should not be presented as a direct human-ground-truth model benchmark.

The pinned implementation distinguishes artifact validation from model loading. `/health/live` returns `status: ok` with a `model_loaded` field; `/health` reports `warming` when the model is not loaded. `/health/ready` returns 503 when artifact validation fails, otherwise calls the model loader and returns the load flag. A loader exception follows the generic error handler and returns 500, so 503 is not the only possible load-failure response.

The documented prediction response includes a label, confidence value, and top-k field for each attribute. A returned confidence is not, by itself, evidence that probabilities are calibrated or that an input is in distribution; those properties need separate evaluation.

The reported benchmark scores are specific to the 13,355-image TextileNet subset, its splits, and the silver-label pipeline. They describe the repository's comparison under that setup; they do not establish performance on other datasets or use cases.

The human label audit and test-set macro-F1 answer different questions. Neither should be collapsed into a single claim of general model accuracy.