AuxSegCount: auxiliary seg-Attention based network for wheat ears counting in field conditions

Accurate wheat ears counting is crucial to the wheat yield estimation. Existing counting methods explore various architectures using density maps for training, and a few works also incorporate an auxiliary network. However, the small wheat ear with background noises makes its detection hard, and the auxiliary network methods tend to ignore interactions with its main network. To mitigate these issues, we propose a novel framework AuxSegCount which includes a segmentation auxiliary network and a main network for wheat ears counting. Unlike density map, the segmentation mask provides more local contexts of wheat ears. We therefore utilise it and introduce the segmentation attention module (SAM) that aims to capture local features around wheat ears. To promote the interactions, we further present the BiAttention Fusion Module (BiFM) that exploits both global information and local contexts into the main network. The experimental results on two datasets show the superiority of our method.