Abstract
Large datasets are required to develop Artificial Intelligence (AI) models in AI powered smart farming for reducing farmers' routine workload, this paper contributes the first large lion-head goose dataset GooseDetect<sup>lion</sup>, which consists of 2,660 images and 98,111 bounding box annotations. The dataset was collected with 6 cameras deployed in a goose farm in Chenghai district of Shantou city, Guangdong province, China. Images sampled from videos collected during July 9 -10 in 2022 were fully annotated by a team of fifty volunteers. Compared with another 6 well known animal datasets in literature, our dataset has higher capacity and density, which provides a challenging detection benchmark for main stream object detectors. Six state-of-the-art object detectors have been selected to be evaluated on the GooseDetect<sup>lion</sup>, which includes one two-stage anchor-based detector, three one-stage anchor-based detectors, as well as two one-stage anchor-free detectors. The results suggest that the one-stage anchor-based detector You Only Look Once version 5 (YOLO v5) achieves the best overall performance in terms of detection precision, model size and inference efficiency.
Introduction
In the world of smart farming, automation is key to reducing the manual workload of monitoring livestock. For the lion-head goose industry—a pillar of the economy in Shantou, China—the stakes are high. These geese are waterfowl and cannot be raised in isolated cages; they live in high-density flocks where diseases like paramyxovirus can spread with devastating speed.
The challenge for AI is that geese in these environments are "visually chaotic." They huddle together, causing massive occlusion (geese blocking each other), and blend into cluttered backgrounds. While datasets exist for cattle, pigs, and even elephants, none capture the specific density and visual profile of lion-head geese. This paper fills that gap by providing a high-quality training ground for computer vision models tailored to poultry.
Methodology and Dataset Construction
The researchers deployed six high-definition cameras at a commercial goose farm, capturing footage from various heights and angles to ensure a multi-perspective view. The farm was divided into several functional zones, including resting, recreation, and nursery areas for both big and small geese.
The primary contribution is the GooseDetect$^{lion}$ dataset, which includes 2,660 images and 98,111 bounding boxes. It is divided into two subsets: GooseDetect$^{lion}_b$ (2,000 images of "big geese" >9 weeks old) and GooseDetect$^{lion}_s$ (660 images of "small geese" 5–8 weeks old). With an average of nearly 37 annotations per image, it is significantly denser than many existing public animal datasets.
The team employed a rigorous three-step annotation pipeline:
- Pre-processing: Removing images where geese were too vague or more than 70% obscured.
- Manual Annotation: A team of 50 volunteers labeled every visible goose with a bounding box and a category tag.
- Refinement: Three experienced mentors reviewed every image to correct mislabeled tags or poorly fitted boxes.
Experimental Results and Analysis
To test the dataset's utility, six representative object detection architectures were evaluated, ranging from "two-stage" detectors like Faster R-CNN to "one-stage" anchor-free models like CenterNet and YOLOv7.
Key Insights
- Density is the Primary Difficulty: The sheer number of geese per frame makes this a "dense object detection" problem. Standard models often struggle when targets overlap significantly.
- Age-Based Performance Gap: All tested models performed better on big geese than small geese. Small geese are smaller targets and tend to cluster more tightly, increasing detection difficulty.
- Deployment Trade-offs: While YOLOv7 achieved the highest Mean Average Precision (mAP) (0.549 for big geese), it is computationally heavy. YOLOv5 emerged as the practical champion for real-world deployment. Despite having significantly fewer parameters (approx. 7 million vs. YOLOv7’s 36 million), it maintained high precision while achieving a blistering 55.78 FPS.
Conclusion
While GooseDetect$^{lion}$ is a major step forward, the authors acknowledge limitations regarding environmental variety and the single-site nature of the data. Future work will focus on night-time monitoring using infrared imagery and expanding to different farm layouts. Additionally, future iterations could include temporal data to support tracking and behavioral analysis, such as identifying sick individuals based on inactivity.
Summary updated