ViT-FFrCnt: A Few-shot Fruit Counting Approach with Multiscale Vision Transformer

Having prior knowledge about the amount of fruit before harvesting can lead to improved planning and decisionmaking in the agriculture sector. In recent years, several detection-based approaches have been proposed to address the problem of fruit counting. However, employing a detection approach for a specific category of fruit requires a large amount of data, which is very costly. Meanwhile, a trained specific class detection model cannot handle new unseen categories without being trained on a large amount of new data. In this paper, we present a novel approach for class-agnostic fruit counting that can count the number of fruits in an arbitrary category with just a few examples. Compared to other fruit detection methods, our few-shot approach only needs a small amount of data to count on a new novel class. Specifically, our proposed model takes advantage of the vision transformer with multi-scale features to improve the accuracy of the counting result. We evaluated our proposed approach in the DeepFruit dataset, which contains six types of fruits (including mango, capsicum, strawberry, apple, avocado, and orange). We compare our method with several stateof-the-art object detection approaches. Our approach archives the MAE of 1.13 and RMSE of 1.48, respectively, on the DeepFruit dataset. Extensive experiments and qualitative results demonstrate the effectiveness of our approach compared to other detection approaches for fruit counting.