Contrastive learning improves phenotyping outcomes in cotton using multi-camera views, suggesting advancements in agricultural technology.
Key Points
The study aims to enhance phenotyping tasks by utilizing a multi-camera dataset to evaluate contrastive learning methods under limited data conditions.
Employs self-supervised learning with synthetic and naturally collected data.
Analyzes the performance of SimCLR and MoCo frameworks for representation learning.
Conducts linear evaluation and semi-supervised learning experiments on cotton boll images.
Evaluates the impact of camera positions and overlaps on detection accuracy.
Achieves a 14% improvement in boll detection mean average precision using multiple camera views.
Identifies optimal camera poses as those with intermediate overlap.
Finds that neither MoCo nor SimCLR consistently outperforms the other.