Abstract
Insects play a crucial role in agricultural production and should not be overlooked. However, there is currently no large-scale dataset available specifically for common insects in orchards. Additionally, datasets for computer vision target detection tasks are limited in the field of insects, which hinders the use of deep learning target detection techniques in orchard insect monitoring. This paper presents the OIDS-45 dataset, which is a large-scale dataset for orchard insect monitoring. The dataset contains 58,585 images of 45 categories of common insects found in orchards. The dataset exhibits a long-tailed distribution, and all images are labeled with borders, making them useful for target detection tasks. The dataset represents the category of orchard insects and has a larger sample size, more categories, and more features in the orchard scenario than previous datasets. We compared our dataset with existing typical insect datasets using advanced target detection algorithms to evaluate its features and quality. The experimental results indicate that current target detection algorithms are not yet capable of accurately identifying and detecting insects in orchards. This is due to the small size of individual insects, the morphological similarities between some species, and the existence of multiple growth stages in some insects. The production and release of this dataset aim to support research in the fields of orchard pest control and insect monitoring in orchards.
Introduction
Orchard management is a high-stakes game where pests can reduce crop yields by 10% and significantly increase production costs. While deep learning has transformed computer vision, its application in orchards remains challenging. Existing insect datasets often focus on broad agricultural categories rather than the specific ecosystem of an orchard and frequently lack the precise bounding box labels required for object detection. Orchard insects present unique technical hurdles: they are often tiny, exhibit morphological similarity to other species, and undergo drastic appearance changes throughout their life cycles. To address these gaps, we introduce OIDS-45, a benchmark dataset featuring over 58,000 labeled images of 45 insect species found in orchard environments.
Methods
The OIDS-45 dataset was developed through a rigorous four-step pipeline:
- Collection: Data was gathered using web crawling and high-resolution field photography/video in orchards across Sichuan Province, China.
- Filtering: Experts and volunteers screened 170,000 candidates to select approximately 58,000 high-quality images.
- Preprocessing & Augmentation: To ensure model robustness, the team applied Gaussian Bilateral Filtering to denoise images while preserving critical edge information.
Further augmentations included Cutout and Mosaic techniques to simulate occlusions, and CosinGAN to generate synthetic training samples for better generalization.
- Expert Annotation: Bounding boxes were verified through three rounds of professional review to ensure taxonomic accuracy.
Dataset Characteristics
OIDS-45 contains 58,585 images across 45 categories. It follows a "long-tailed" distribution, mirroring actual ecological distributions where a few species are highly prevalent while many others are rare.
A key differentiator of OIDS-45 is its inclusion of 14 categories of beneficial insects, such as ladybirds and lacewings. This supports "ecological farming" by allowing monitors to track natural predators alongside pests. Compared to the IP102 dataset, OIDS-45 offers more specialized localization data, providing a stronger training signal for detection tasks.
Key Insights and Results
The researchers benchmarked 12 detection models, including YOLOv8, Faster R-CNN, and CenterNet.
- Top Performers: ViT-FRCNN and CenterNet achieved the highest accuracy, with a Mean Average Precision (mAP) of approximately 81.3%.
- The Life Cycle Trap: Models struggled with insects that look different across life stages (e.g., larvae vs. adults), with accuracy dropping to around 72% for multi-stage species.
- Small Object Bottleneck: Tiny insects remain difficult to detect as feature maps often lose critical information during the downsampling process of deep neural networks.
- Benchmark Comparison: Models generally performed better on OIDS-45 than on the detection-labeled subset of IP102. For instance, YOLOv7 reached 80.12% mAP on OIDS-45 versus 67.71% on IP102.
Conclusion and Future Work
OIDS-45 provides a robust foundation for orchard insect monitoring, though the long-tail effect remains a challenge for learning rare species features. Future research will focus on incorporating multimodal data (such as insect sounds or infrared imagery), utilizing reinforcement learning for more efficient annotation, and exploring federated learning to train models across diverse geographic locations without centralizing sensitive data.
Summary updated