Tomatod: Evaluation of Object Detection Algorithms on a New Real-World Tomato Dataset

Abstract

Abstract. The integration of modern technologies in farming poses a challenging task to the research community. In this work, the task of selective cropping and treating is considered, whereas learning algorithms can provide essential assistance on crop growth and disease prediction, species recognition and fruit detection. In this paper, we introduce a highly specialized object detection (OD) and classification dataset of tomato fruits that contains class information for the ripening stage of each tomato fruit apart from its corresponding bounding box. With this dataset we aim to encourage the development of task-specific production ready object detection algorithms, as well as to evaluate and provide a baseline result of common state-of-the-art generic OD algorithms. In detail, a thorough presentation of the most common OD datasets takes place, where we discuss both generic OD and some highly specialized datasets. Our dataset contains more than 250 images and 2400 annotations in total. The dataset contains class information for three ripening stages of a tomato fruit provided by expert agriculturists, while providing views consistent with the targeted real-world use case scenario. Compared to other OD datasets our proposition differs in core areas such as the quality of the annotations, the object size distribution and the public availability. Evaluating the performance in our dataset for six object detection models we draw conclusions about the strength and weaknesses of each one’s performance. Finally, we present a future roadmap of revisions and discuss some future research topics that could improve the performance of OD algorithms in our dataset.

Scale-Awareness for More Accurate Object Detection Using Modified Single Shot Detectors

“…The scale-aware approach presented in this work constitutes a generic technique to adapt mainstream single-shot detectors so that they can take advantage of the singular object scale distribution of the target training and testing dataset. Here, we present the novel detector adaptation pipeline in a more streamlined fashion, since we focus on the adaptation of SSD Mo-bileNet v2 for the task-specific TomatOD dataset (Tsironis et al, 2020); however, our technique can expand to literally any single-shot detector and dataset.…”

“…We validate our proposition on "small object detec- * Corresponding author tion" which is a research sub-field of particular interest due to its inherent "by-definition" difference in object size distribution compared to standard OD datasets. The evaluation is based on the task-specific but intriguing TomatOD dataset (Tsironis et al, 2020). This work introduces two major contributions: firstly, we present a straightforward pipeline to configure an object detection model so that it can perform optimally in the size range of a given target object; secondly, we introduce a novel training strategy which significantly improves the detection performance of a model avoiding the complexity of other similar methods.…”

Benchmark of Deep Learning and a Proposed HSV Colour Space Models for the Detection and Classification of Greenhouse Tomato

“…There are already some studies that compare and evaluate different SSD and YOLO architectures, with different CNNs, especially when it comes to tomato detection [28,49,50,55,57]. However, the only and most relevant paper found on tomato classification based on one-stage detectors was the one by Mutha et al [58], who compared a CNN with a YOLO model to classify tomatoes into 3 classes (unriped, riped and damaged).…”

Advanced Technology in Agriculture Industry by Implementing Image Annotation Technique and Deep Learning Approach: A Review

“…Thus, annotated images for training supervised deep learning models achieve acceptable performance levels. Most of the studies applied supervised learning, as this method promises high accuracy as proposed in [38][39][40]. Another attractive annotation method is based on unsupervised learning.…”