Application of amodal segmentation for shape reconstruction and occlusion recovery in occluded tomatoes

Common object detection and image segmentation methods are unable to accurately estimate the shape of the occluded fruit. Monitoring the growth status of shaded crops in a specific environment is challenging, and certain studies related to crop harvesting and pest detection are constrained by the natural shadow conditions. Amodal segmentation can focus on the occluded part of the fruit and complete the overall shape of the fruit. We proposed a Transformer-based amodal segmentation algorithm to infer the amodal shape of occluded tomatoes. Considering the high cost of amodal annotation, we only needed modal dataset to train the model. The dataset was taken from two greenhouses on the farm and contains rich occlusion information. We introduced boundary estimation in the hourglass structured network to provide a priori information about the completion of the amodal shapes, and reconstructed the occluded objects using a GAN network (with discriminator) and GAN loss. The model in this study showed accuracy, with average pairwise accuracy of 96.07%, mean intersection-over-union (mIoU) of 94.13% and invisible mIoU of 57.79%. We also examined the quality of pseudo-amodal annotations generated by our proposed model using Mask R-CNN. Its average precision (AP) and average precision with intersection over union (IoU) 0.5 (AP50) reached 63.91%,86.91% respectively. This method accurately and rationally achieves the shape of occluded tomatoes, saving the cost of manual annotation, and is able to deal with the boundary information of occlusion while decoupling the relationship of occluded objects from each other. Future work considers how to complete the amodal segmentation task without overly relying on the occlusion order and the quality of the modal mask, thus promising applications to provide technical support for the advancement of ecological monitoring techniques and ecological cultivation.