BBoxCut: A Targeted Data Augmentation Technique for Enhancing Wheat Head Detection Under Occlusions Journal: Computers and Electronics in Agriculture
Yasashwini Sai Gowri P, Karthik Seemakurthy, Andrews Agyemang Opoku, Sita Devi Bharatula · 2025
Read BBoxCut: A Targeted Data Augmentation Technique for Enhancing Wheat Head Detection Under Occlusions Journal: Computers and Electronics in Agriculture by Yasashwini Sai Gowri P, Karthik Seemakurthy, Andrews Agyemang Opoku, Sita Devi Bharatula (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-24032.txt.
Introduction
Yasashwini Sai Gowri P and three co-authors improve wheat head detection in field photos, in this 2025 paper. Breeders count and measure wheat heads to pick lines for crossing, but hand work is slow. Camera systems struggle when leaves hide heads, heads overlap, light shifts, or motion blurs the shot.
Their trick, BBoxCut, trains detectors by masking random local patches of training images, faking the very blockages that fool them. The masks sit inside the labeled boxes around wheat heads. Three modern detectors were trained with and without the trick. Mean accuracy rose by 2.76, 3.26, and 1.9 points across the three. Gains were largest on hidden heads, exactly the hard cases. The authors offer BBoxCut as a cheap add-on for crop imaging pipelines.
Catalog record
| CREATOR | Yasashwini Sai Gowri P, Karthik Seemakurthy, Andrews Agyemang Opoku, Sita Devi Bharatula |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 30 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-24032
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
BBoxCut: A Targeted Data Augmentation Technique for Enhancing Wheat Head Detection Under Occlusions Journal: Computers and Electronics in Agriculture Yasashwini Sai Gowri P, Karthik Seemakurthy, Andrews Agyemang Opoku, Sita Devi Bharatula Abstract Wheat plays a critical role in global food security, making it one of the most extensively studied crops. Accurate identification and measurement of key characteristics of wheat heads are essential for breeders to select varieties for cross-breeding, with the goal of developing nutrient-dense, resilient, and sustainable cultivars. Traditionally, these measurements are performed manually, which is both time-consuming and inefficient. Advances in digital technologies have paved the way for automating this process. However, field conditions pose significant challenges, such as occlusions of leaves, overlapping wheat heads, varying lighting conditions, and motion blur. In this paper, we propose a novel data augmentation technique, BBoxCut, which uses random localized masking to simulate occlusions caused by leaves and neighboring wheat heads. We evaluated our approach using three state-of-the-art object detectors and observed mean average precision (mAP) gains of 2.76, 3.26, and 1.9 for Faster R-CNN, FCOS, and DETR, respectively. Our augmentation technique led to significant improvements both qualitatively and quantitatively. In particular, the improvements were particularly evident in scenarios involving occluded wheat heads, demonstrating the robustness of our method in challenging field conditions. † † graphicalabstract: † † highlights: Proposed a random localised masking based data augmentation technique for improved occluded wheat detection. Improved detection performance for two scenarios: a) wheat heads that overlap each other and b) wheat heads that are occluded by leaves. The effectiveness of the proposed approach was evaluated across three different object detection architectures. 1 Introduction Wheat is a staple crop for more than three billion people around the world and serves as a critical source of essential nutrients. To meet the increasing demands of a growing global population, agricultural production must double by 2050 [ 1 ] . Given its status as the most widely grown crop, accurate monitoring of wheat yield is crucial to ensure global food security [ 2 ] . Estimating wheat yield and making informed decisions about which wheat varieties to cross-breed for more nutritious, resilient, and high-yield cultivars are dependent on accurate detection of wheat head. Traditionally, yield prediction has relied on expert forecasts and manual counting of wheat heads, which are both time consuming and prone to inaccuracies [ 3 ] . The Global Wheat Head Detection (GWHD) 2021 dataset [ 4 ] was introduced to facilitate the development of automated systems capable of measuring the visual traits of wheat heads. Despite significant advances in object detection models [ 5 ] , detecting wheat heads in unconstrained real-world field conditions presents considerable challenges, as illustrated in Fig. 1 . Fig. 1 (a) shows occlusions caused by leaves, while Fig. 1 (b) highlights wheat heads that overlap one another. In both cases, state-of-the-art object detectors often struggle to accurately locate wheat heads. This performance degradation is likely due to the lack of diverse training examples that effectively represent occlusion scenarios. Furthermore, Fig. 2 reveals that the test domains exhibit higher wheat head densities compared to the training set, further increasing the likelihood of occlusions caused by leaves and neighboring wheat heads, as shown in Fig. 1 . Such occlusions are expected to significantly affect the average performance of detectors across unseen test domains. (a) (b) Figure 1: Samples from GWHD 2021 dataset. (a) Wheat heads partially occluded by leaves, (b) Wheat heads occluding each other. Occlusions are indicated by the red circles. We propose BBoxCut, a data augmentation technique that applies random localized masking to simulate realistic occlusions in wheat fields. Unlike Cutout [ 6 ] , which randomly masks square regions throughout the image, BBoxCut specifically masks portions of the bounding boxes, mimicking partial occlusions by leaves or neighboring wheat heads. By incorporating BBoxCut into the data augmentation pipeline for the training of object detection models (Faster R-CNN, FCOS, and DETR) we observed significant improvements in mean Average Precision (mAP), particularly under occlusion scenarios. This enhancement is crucial to accurately predict wheat yields, directly impacting farmer profitability. The following is a list of our contributions proposed in this paper. • We proposed a new data augmentation technique that improves wheat head localization in the presence of occlusions. • We propose using a histogram-based dominant color estimation for mask generation. • We have shown the validity of our approach on three different types of detectors such as Faster R-CNN, FCOS, and DETR. • Our comparison with state-of-the-art masking-based data augmentation approaches demonstrates improved detection capability using the proposed technique. Figure 2: Average wheat head count per image in GWHD 2021 dataset. 2 Related work In this section, we provide an overview of related works, taking a broad look at deep learning-based object detection, recent efforts to improve wheat detection accuracy, and studies that explicitly model occlusion in the context of object detection. We also highlight how our approach differs from existing techniques. 2.1 Deep Learning for object detection Deep neural networks have been adopted for various computer vision tasks, such as object detection, due to their exceptional performance in feature extraction [ 7 ] . Most object detection networks can be classified into two categories: single-stage [ 8 , 9 , 10 , 11 , 12 , 13 ] and two-stage [ 14 , 15 ] . Although most existing approaches use convolutional neural network (CNN) architectures, transformer-based architectures [ 16 , 17 ] have been gaining importance. Despite significant improvements in speed and accuracy, deep learning-based object detection networks generally do not explicitly model occlusion [ 18 ] , which is common in many real-world scenarios [ 4 , 19 ] . 2.2 Wheat Head Detection The importance of wheat for global food security has driven researchers to apply various detection techniques to detect wheat heads. Earlier work used image processing and machine learning techniques for wheat head detection. [ 20 ] used color-based features to count wheat heads. However, wheat ears, leaves, and stalks in the wheat field share similar color and texture characteristics. In addition, as the wheat plant matures, the color of different parts of the plant changes. As a result, using color cues to reliably identify wheat ears proved challenging. [ 21 ] successfully detected and counted wheat ears using the support vector machine approach. [ 22 ] utilized the k-means approach to partition the wheat head and obtain identification. To segment and count wheat heads, the twin support vector machine segmentation approach was used in [ 23 ] . [ 24 ] segmented wheat heads from backgrounds using Fourier filtering and Fourier transform. Due to the limitations of traditional object detection methods, researchers in recent years have adopted deep learning-based wheat head localization techniques. [ 25 ] used R-CNN to detect and analyze wheat heads. Similarly, [ 26 ] applied ablation in color, affine variation, and random cutout on the Global Wheat Head Detection (GWHD) data set and trained the model on a pre-trained YOLOv5 model. Although the performance of the model was not overly impressive, the random cutout showed an increase in performance compared to the ablation of color and the affine variation. [ 27 ] successfully trained a ResNet-50 model to detect and