Background-invariant re-identification of dogs from camera-trap videos in non-controlled environments
Résumé
This paper addresses the general problem of re-identification in natural conditions with multiple camera traps, poor video quality and small datasets. We focus on generalizable re-identification of dogs in cross-camera setups, adapting from short-term to long-term scenarios. Long-term re-identification across multiple cameras presents challenges due to variations in background, camera angles, and lighting conditions. While realistic, few animal re-identification methods are tested under such settings, mainly due to the lack of datasets and high complexity of annotation. Short-term datasets are often used to train re-identification networks, since they can be simply generated through web scraping algorithms. We introduce two publicly available datasets: the YT-BB-Dog, a short-term dataset with 2723 dogs from YouTube videos, and the Sibetan, a long-term dataset featuring 59 dogs recorded over 5 days and 12 cameras placed in Sibetan, Bali, Indonesia. Our goal is to use the YT-BB-Dog to train a feature extractor robust to covariate shifts, enabling better generalization in unknown domains. Our experiments revealed that state-of-the-art (SOTA) methods trained on the YT-BB-Dog are heavily influenced by background variations and perform poorly on complex scenarios like Sibetan. To address this, we propose Background Invariant Feature extractOR (BIFOR), a three-step method that leverages a novel mini-batch sampling technique with triplet loss and online hard mining. BIFOR achieves SOTA performance on Sibetan, improving rank-1 accuracy of the baseline by more than 9%. We also present a complete pipeline combining detection, tracking, and re-identification based on BIFOR.
| Origine | Fichiers éditeurs autorisés sur une archive ouverte |
|---|---|
| Licence |