Open-Set Distance Cutoff

Overview

In open-set recognition, test images may belong to classes not seen during training (e.g., "new_whale"). After retrieving k nearest neighbors by embedding distance, if all distances exceed a learned cutoff, insert the unknown class into the prediction list. The cutoff is tuned on a validation set to maximize the ranking metric (e.g., MAP@5).

Quick Start

import numpy as np

def predict_with_unknown(query_dists, query_nbs, train_labels,
                         unknown_label='new_whale', dcut=3.8, top_k=5):
    predictions = []
    for i in range(len(query_dists)):
        seen = {}
        for j in range(query_nbs.shape[1]):
            label = train_labels[query_nbs[i, j]]
            dist = query_dists[i, j]
            if dist > dcut and unknown_label not in seen:
                seen[unknown_label] = dcut
            if label not in seen:
                seen[label] = dist
            if len(seen) >= top_k:
                break
        preds = sorted(seen.items(), key=lambda x: x[1])[:top_k]
        predictions.append([p[0] for p in preds])
    return predictions

Workflow

1. Compute embeddings for train and test sets
2. Find k nearest neighbors for each test sample
3. For each test query, iterate neighbors by distance
4. Insert unknown class when distance exceeds cutoff dcut
5. Tune dcut on validation set to maximize MAP@k or accuracy

Key Decisions

• Cutoff tuning: sweep dcut values on validation set; optimal value depends on embedding space scale
• Insertion point: unknown class enters at the position where distance first exceeds cutoff
• Distance metric: Euclidean on L2-normalized embeddings ≈ cosine distance
• When to use: any retrieval task where query may not match any gallery class

References

• Similarity DenseNet121 [0.805LB]