STP-AgriData: Pomelo Detection Benchmark

• STP-AgriData is a specialized single-class dataset designed for precise Shatian pomelo detection in complex orchard scenes.
• It fuses field-collected and public images to tackle challenges such as multi-scale variation, occlusion, and complex lighting conditions.
• The dataset serves as a benchmark for advanced detectors like SDE-DET to enable applications in robotic harvesting and yield estimation.

Searching arXiv for the primary STP-AgriData paper and a few related agricultural data/platform references for context. STP-AgriData is a custom Shatian pomelo dataset constructed for single-class object detection in real orchard conditions, designed to support pomelo localization, automated robotic harvesting, and downstream tasks such as maturity analysis and yield estimation. It combines field-collected images from an actual Shatian pomelo orchard with publicly sourced online images, and is described as, to the authors’ knowledge, the first Shatian pomelo dataset that fuses field-collected and public data in order to increase diversity and comprehensiveness (Hu et al., 24 Sep 2025). Within the associated detection study, STP-AgriData functions not merely as a data repository but as the central benchmark that encodes multi-scale variation, occlusion, small-object difficulty, background similarity, and complex lighting—precisely the orchard conditions that motivate the SDE-DET detector and define its evaluation regime (Hu et al., 24 Sep 2025).

1. Definition, scope, and targeted visual difficulties

STP-AgriData is a single-class object detection dataset whose sole class is “Shatian pomelo.” Its stated purpose is to support detection in complex orchard scenes relevant to localization in the canopy, automated robotic harvesting, and, once fruits are reliably detected and counted, maturity analysis and yield estimation (Hu et al., 24 Sep 2025). The dataset is therefore specialized rather than generic: it does not target general fruit detection, nor does it provide multi-class labels across fruit types.

The dataset is explicitly built to represent the characteristic visual difficulties of Shatian pomelo orchards. These include strong multi-scale variation, since pomelos close to the camera appear large while distant pomelos become small targets in the same image; frequent occlusion by leaves, trunks, and other fruits; dense clusters that create overlapping bounding boxes; background similarity, because pomelo color and texture are similar to leaves and trunks; and complex lighting, especially overexposure and dark shadows under strong sun and high contrast (Hu et al., 24 Sep 2025). Small-object detection is particularly central, because distant pomelos occupy very few pixels and exhibit heavily degraded features.

These properties distinguish STP-AgriData from more general agricultural image collections. The source paper notes that existing fruit-detection datasets covering apples, citrus, peppers, strawberries, cherry tomatoes, camellia fruits, and winter jujube do not focus on Shatian pomelo and do not fully capture the combination of multi-scale variation, heavy occlusion, background similarity, and small distant fruits that characterizes Shatian pomelo orchards (Hu et al., 24 Sep 2025). This suggests that the dataset is best understood as a task-specific benchmark for precision pomelo detection rather than a broad-spectrum agricultural vision corpus.

2. Data acquisition, site conditions, and corpus composition

Field data in STP-AgriData were collected in Taipingshan Village, Ganzhou City, Jiangxi Province, China, at coordinates 114°51′45″ E, 25°20′52″ N, with elevation 193.1 m and orchard area about 49.45 mu (approximately 3.3 ha) (Hu et al., 24 Sep 2025). Collection was performed under natural light only, on sunny days rather than rain, fog, or overcast conditions, and images were mainly captured in the afternoon to avoid harsh midday overexposure and low light at dawn or dusk. The sampling period was winter, from enlargement to maturity, corresponding to a harvest-relevant stage for Shatian pomelo in that region (Hu et al., 24 Sep 2025).

Tree sampling and imaging were controlled to standardize scene geometry. Only mature trees aged 5–7 years, with stable fruiting performance, were used. Trees were about 2 m high; images were captured from the outer south side of the crown; and agronomic management was uniform, including organic fertilizer, integrated water–fertilizer irrigation, and no special treatments such as extra pesticide or pruning (Hu et al., 24 Sep 2025). Imaging hardware consisted primarily of Hikvision DS-8632N‑I8 cameras producing RGB JPEG images at 1920 × 1080 resolution, with a camera–tree horizontal distance of 3–5 m, camera height of 2 m, and factory-default aperture, shutter, and ISO settings (Hu et al., 24 Sep 2025).

To increase diversity, the dataset also includes Shatian pomelo images collected from public online sources. These add a wider variety of regions, orchard layouts, backgrounds, camera viewpoints, and conditions, and are stated to improve generalization by exposing the detector to non-local orchard scenes (Hu et al., 24 Sep 2025). The combination of in-situ orchard imagery with public imagery is one of the dataset’s defining features.

The paper reports 317 original images after cleaning, with one damaged image removed before processing. The train–test split is 60%/40%, with 190 images for training and 127 for testing, using a fixed random seed for reproducibility. Six augmentations are applied to the 190 training images—brightness adjustment, contrast adjustment, denoising, grayscale transformation, horizontal flipping, and vertical flipping—bringing the total number of images to 1,330. All images are resized to 640 × 640 for training (Hu et al., 24 Sep 2025).

Component	Value	Notes
Original images	317	One damaged image removed before processing
Train / test split	190 / 127	60% / 40%, fixed random seed
Augmentations	6	Brightness, contrast, denoising, grayscale, horizontal flip, vertical flip
Total after augmentation	1,330	Used in training SDE-DET and baselines
Training input size	640 × 640	All images resized before model input

The acquisition strategy reflects a narrow but realistic observational regime: ground-based RGB imagery under natural orchard conditions. A plausible implication is that STP-AgriData is especially appropriate for fixed-camera or field-robot perception pipelines, rather than for aerial or multispectral analysis.

3. Annotation structure and benchmark character

The annotation protocol is only partially described, but the paper makes clear that STP-AgriData is a 2D bounding-box detection dataset with a single class, “Shatian pomelo,” and that each image has a corresponding annotation file in a unified directory structure ensuring one-to-one mapping (Hu et al., 24 Sep 2025). There is no mention of instance segmentation masks, keypoints, 3D information, maturity labels, occlusion categories, difficulty levels, inter-annotator agreement, or dedicated quality-control procedures.

Accordingly, the dataset is tailored to 2D single-class detection under complex visual conditions rather than to broader multimodal or multi-task supervision. This should not be conflated with a phenotyping dataset or a crop-monitoring dataset in the remote-sensing sense. The label space is deliberately minimal, and its complexity lies in scene conditions rather than semantic breadth.

As a benchmark, STP-AgriData is tied closely to the evaluation of SDE-DET and a series of detector baselines. All models are trained and tested under the same hardware and software conditions and on the same splits, which the source paper presents as enabling fair comparison (Hu et al., 24 Sep 2025). Standard object detection metrics are used:

$\text{Precision} = \frac{TP}{TP + FP} \tag{10}$

$\text{Recall} = \frac{TP}{TP + FN} \tag{11}$

$AP = \int p(r)\,dr \tag{12}$

$mAP = \frac{1}{N} \sum_{i=1}^{N} AP_i \tag{13}$

$$IoU = \frac{\text{Area of Intersection}}{\text{Area of Union}} \tag{14}$$

$$F_1 = 2 \cdot \frac{\text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}} \tag{15}$$

Since the dataset is single-class, $N=1$, though the paper still reports mAP at IoU threshold 0.5 and COCO-style [email protected]:0.95 (Hu et al., 24 Sep 2025).

4. Architectural co-design with SDE-DET

STP-AgriData is unusual in that the dataset’s visual structure directly drives detector design. The source paper states that SDE-DET is built specifically around the challenges encoded in STP-AgriData: multi-scale variation, occlusion, and small-object detection (Hu et al., 24 Sep 2025). The model is an improved YOLOv8n-based detector whose three principal modifications correspond to those dataset properties.

First, the model introduces the Star Block into the early backbone to preserve high-dimensional information without large computational overhead. This is motivated by the claim that the original YOLOv8n backbone exhibits a large dimensional gap between the input image and the first convolutional layer, causing early feature loss for small distant pomelos and texture-similar backgrounds (Hu et al., 24 Sep 2025). The paper gives the Star operation as

$\sum_{i=1}^{d+1} \sum_{j=1}^{d+1} y_i y_j \tag{1}$

or equivalently

$$\sum_{i=1}^{d+1} \sum_{j=i}^{d+1} y(i,j) \, y_i y_j \tag{2}$$

with

$$y(i,j) = \begin{cases} w_1 w_1, & i = j \ w_1 w_2 + w_1 w_2, & i \neq j \end{cases} \tag{3}$$

and $\text{Recall} = \frac{TP}{TP + FN} \tag{11}$0, where $\text{Recall} = \frac{TP}{TP + FN} \tag{11}$1 is the number of input channels (Hu et al., 24 Sep 2025). Architecturally, an early 16-channel convolutional block is replaced by a Conv block with 8 channels, followed by Star Block, followed by a Conv block with 32 channels (Hu et al., 24 Sep 2025).

Second, SDE-DET integrates Deformable Attention after the SPPF layer in the backbone to address occlusion. The paper introduces standard attention with

$\text{Recall} = \frac{TP}{TP + FN} \tag{11}$2

$\text{Recall} = \frac{TP}{TP + FN} \tag{11}$3

$\text{Recall} = \frac{TP}{TP + FN} \tag{11}$4

$\text{Recall} = \frac{TP}{TP + FN} \tag{11}$5

and then describes deformable attention as learning offsets for sampling positions:

$\text{Recall} = \frac{TP}{TP + FN} \tag{11}$6

using bilinear interpolation at deformed positions $\text{Recall} = \frac{TP}{TP + FN} \tag{11}$7 (Hu et al., 24 Sep 2025). The design rationale is adaptive focus on occluded or partially visible pomelo regions.

Third, Efficient Multi-Scale Attention (EMA) is inserted in the neck to handle multi-scale variation and small targets. The paper describes channel grouping $\text{Recall} = \frac{TP}{TP + FN} \tag{11}$8 with each $\text{Recall} = \frac{TP}{TP + FN} \tag{11}$9, and 2D global average pooling per channel:

$AP = \int p(r)\,dr \tag{12}$0

EMA is said to preserve channel information, produce refined pixel-level attention on high-level feature maps, and improve efficient multi-scale fusion for small, distant pomelos (Hu et al., 24 Sep 2025).

This tight dataset–model coupling is methodologically significant. It indicates that STP-AgriData is not only a benchmark but also a design specification for detector architecture under orchard constraints.

5. Evaluation results and diagnostic findings

On the STP-AgriData test set, SDE-DET achieves Precision 0.883, Recall 0.771, [email protected] 0.838, [email protected]:0.95 0.497, and F1-score 0.823 (Hu et al., 24 Sep 2025). Compared with YOLOv8n, whose scores are Precision 0.862, Recall 0.700, [email protected] 0.787, [email protected]:0.95 0.458, and F1 0.773, the gains are +2.1 percentage points in Precision, +7.1 in Recall, +5.1 in [email protected], +3.9 in [email protected]:0.95, and +5.0 in F1 (Hu et al., 24 Sep 2025).

Against other YOLO variants, SDE-DET improves on [email protected] by +2.9 over YOLOv8s, +3.4 over YOLOv9c, +7.3 over YOLOv10n, and +5.9 over YOLOv10s. Against mainstream detectors, Faster R-CNN achieves 0.581 / 0.280 for [email protected] / [email protected]:0.95, Cascade R-CNN 0.618 / 0.359, CenterNet 0.681 / 0.373, RTMDet-m 0.749 / 0.430, DINO-4scale 0.717 / 0.400, DDQ-4scale 0.740 / 0.414, and RT-DETR 0.787 / 0.459, whereas SDE-DET reaches 0.838 / 0.497 (Hu et al., 24 Sep 2025). The model is also lightweight, with 3.29M parameters, 32.4 GFLOPS, and model size about 6.69 MB (Hu et al., 24 Sep 2025).

Ablation on STP-AgriData isolates the effects of the three architectural modifications. Starting from the YOLOv8n baseline, Deformable Attention alone yields Precision 0.870, Recall 0.716, [email protected] 0.794, [email protected]:0.95 0.456, and F1 0.786; adding Star Block gives 0.875, 0.738, 0.817, 0.487, and 0.800; adding EMA produces the final SDE-DET scores of 0.883, 0.771, 0.838, 0.497, and 0.823 (Hu et al., 24 Sep 2025). The paper interprets these increments as evidence that Deformable Attention improves occlusion handling, Star Block improves early feature preservation, and EMA yields the largest gains in recall and overall mAP.

Qualitative analysis identifies the main remaining failure modes: small distant pomelos in upper image regions, heavy sunlight, dense clusters with overlap, and background confusion from leaves with pomelo-like color (Hu et al., 24 Sep 2025). Grad-CAM heatmaps show stronger emphasis on fruit edges and relevant color regions for SDE-DET than for the baseline. The paper also notes that some missed and false detections remain, which underscores that STP-AgriData is a difficult dataset rather than a saturated one.

A common misconception would be to interpret the reported state-of-the-art result as evidence that Shatian pomelo detection is effectively solved. The source material does not support that interpretation. Instead, it stresses persistent false positives and missed detections under challenging lighting, scale, and occlusion conditions (Hu et al., 24 Sep 2025).

6. Position within agricultural data infrastructures and related benchmarks

STP-AgriData is a narrowly specialized orchard-vision dataset, and that specialization differentiates it from broader agricultural data infrastructures. FAIR-oriented agricultural platforms such as ADMA emphasize cross-disciplinary semantic data management, portals, APIs, HPC-enabled analysis, and provenance across genomics, phenomics, environmental, agronomic, and socioeconomic data (Pan et al., 2023). Harmonized spatio-temporal resources such as SCARFACE integrate climate, air quality, emissions, land cover, farm techno-economics, livestock, and socio-economic indicators into annual panels over agrarian sub-regions (Maranzano et al., 16 Apr 2026). Remote-sensing benchmarks such as AgriPotential provide multi-spectral, multi-temporal Sentinel‑2 patches with pixel-level agricultural potential labels for several crop types and ordinal classes (Sakka et al., 13 Jun 2025).

By contrast, STP-AgriData is not a FAIR data management infrastructure, not a multi-domain panel, and not a remote-sensing benchmark for crop suitability. It is a ground-based RGB orchard dataset for 2D single-class detection. Its realism comes from orchard scene complexity rather than from thematic breadth. This distinction matters because “AgriData” in the name should not be mistaken for a general-purpose agricultural data platform.

At the same time, STP-AgriData shares several design principles with these broader efforts. Like AgriPotential, it encodes task-relevant spatial variation and supports model benchmarking under explicitly difficult conditions (Sakka et al., 13 Jun 2025). Like ADMA, it would plausibly benefit from more explicit metadata, provenance, and interoperable access mechanisms if it were to evolve into a reusable community resource (Pan et al., 2023). Like SCARFACE, it illustrates the value of tailoring the data representation to the agricultural unit of interest—in this case, canopy-level visual scenes rather than agrarian sub-regions (Maranzano et al., 16 Apr 2026).

7. Availability, practical use, and prospective extensions

The paper states that the data “will be made available upon request,” implying that STP-AgriData is not openly hosted via a public URL in the paper and that licensing details are not specified (Hu et al., 24 Sep 2025). By contrast, the SDE-DET source code is provided at https://github.com/mistletoe111/SDE-DET (Hu et al., 24 Sep 2025). Recommended usage follows the paper’s preprocessing: align image–annotation pairs in a unified directory, clean damaged images, resize to 640 × 640, optionally apply the same six augmentations, and use the same 6:4 train/test split or explicitly report deviations (Hu et al., 24 Sep 2025).

The immediate practical application is automated robotic harvesting of Shatian pomelo, for which detection and localization are prerequisite operations. The paper also identifies maturity analysis and yield estimation as downstream tasks enabled by reliable detection and counting (Hu et al., 24 Sep 2025). Because SDE-DET is lightweight, with model size around 6.69 MB, the authors regard it as suitable for embedded systems on harvest robots (Hu et al., 24 Sep 2025).

Several limitations remain structural. STP-AgriData does not include segmentation masks, keypoints, maturity labels, occlusion tags, or 3D information. It is collected under sunny natural-light conditions rather than across a wider range of weather regimes. The annotation process is not documented in terms of inter-annotator agreement or formal quality control (Hu et al., 24 Sep 2025). This suggests several plausible extensions: multi-task annotation for maturity and occlusion, multi-view or depth augmentation for robotic grasping, and broader acquisition under rain, fog, overcast, or nocturnal conditions. Such extensions, however, are prospective; they are not part of the present dataset specification.

In its current form, STP-AgriData occupies a specific niche in agricultural computer vision: a Shatian pomelo benchmark for precision detection in complex orchard environments, constructed to reflect realistic harvesting conditions and to stress-test lightweight detectors under severe occlusion, clutter, scale variation, and illumination imbalance (Hu et al., 24 Sep 2025).