A Novel Active Learning Approach for Improving Classification of Unlabeled Video Based on Deep Learning Techniques

Résumé

Abstract Deep learning has made great strides in computer vision, but requires substantial labeled data. Manual video annotation is extremely laborious, as annotators must watch entire videos to provide accurate labels. Active learning reduces this effort by selectively querying the most informative examples for labeling. We introduce a unique active learning framework designed specifically for enhancing the efficiency of video classification. It minimizes annotator effort by identifying representative and informative frames, rather than full videos. Our batch selection approach pinpoints useful videos using uncertainty and diversity criteria. Representative sampling then extracts exemplar frames from each video. Annotators only need to label these frames, avoiding watching full videos. Experiments demonstrate our method successfully reduces human annotation cost for video classification. By querying informative, representative frames, it acquires crucial labels without needing exhaustive manual review of all data. We introduce an active learning technique specifically for alleviating the video annotation burden. By requesting labels for a small set of exemplar frames, it achieves high accuracy while minimizing human effort. This makes large-scale video classification more feasible.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Voir l'article sur le site de la revue

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0