Weakly Supervised Underwater Fish Segmentation using Affinity LCFCN

Estimating fish body measurements like length, width, and mass has potential applications in marine and aquaculture productivity. The paper proposes a segmentation model requiring only point-level supervision (single click per fish, ~1 second per fish) rather than per-pixel labels (up to 2 minutes per fish). The method uses a fully convolutional neural network with two branches: one outputs per-pixel scores, another outputs an affinity matrix. These are combined using random walk refinement and trained end-to-end with LCFCN loss, termed Affinity-LCFCN (A-LCFCN). Experiments on the DeepFish dataset show A-LCFCN outperforms fully-supervised models at fixed annotation budgets and achieves better results than standard baselines.