Evaluation of model generalization for growing plants using conditional learning

This paper aims to solve the lack of generalization of existing semantic segmentation models in the crop and weed segmentation domain. We compare two training mechanisms, classical and adversarial, to understand which scheme works best for a particular encoder-decoder model. We use simple U-Net, SegNet, and DeepLabv3+ with ResNet-50 backbone as segmentation networks. The models are trained with cross-entropy loss for classical and PatchGAN loss for adversarial training. By adopting the Conditional Generative Adversarial Network (CGAN) hierarchical settings, we penalize different Generators (G) using PatchGAN Discriminator (D) and L1 loss to generate segmentation output. The generalization is to exhibit fewer failures and perform comparably for growing plants with different data distributions. We utilize the images from four different stages of sugar beet. We divide the data so that the full-grown stage is used for training, whereas earlier stages are entirely dedicated to testing the model. We conclude that U-Net trained in adversarial settings is more robust to changes in the dataset. The adversarially trained U-Net reports 10% overall improvement in the results with mIOU scores of 0.34, 0.55, 0.75, and 0.85 for four different growth stages.