a two-stage pipeline, still very much a work in progress

Catnip tries to answer one important question: is Izutsumi in this image? It runs a two-stage pipeline: first detect, then identify over Dungeon Meshi manga pages. Nothing is exported yet.

The pipeline

1. Localization (SAHI + YOLO26n). Find where bodies and faces are. A YOLO26n model runs over full manga pages that SAHI slices into overlapping 640×640 patches to handle different scales. The training set is pre-sliced with the same parameters.

2. Re-Identification (refactor/reID branch). Find Izutsumi herself. Metric learning:

  • Backbone is a ResNet18 (ImageNet-pretrained) with GeM pooling.
  • Loss* Triplet Loss with hard-negative mining.
  • Manually labeled crops Label Studio crops plus Stage 1 detector crops.
  • FAISS IndexFlatIP over L2-normalized embeddings w/ cosine similarity for matching. Optional LogisticRegression re-ranking.

Datasets

The same annotations feed both stages:

Updated: