---
title: Multi-Crop3D
canonical_url: "https://www.modelscope.cn/datasets/cau721582/Multi-Crop3D"
md_url: "https://www.modelscope.cn/datasets/cau721582/Multi-Crop3D.md"
repository: cau721582/Multi-Crop3D
chinese_name: "多作物3D点云数据集"
last_updated: 2026-05-07
license: "Apache License 2.0"
storage_size: "3.3 GB"
downloads: 228
stars: 0
---

# Multi-Crop3D

> Multi-Crop3D - cau721582 在 ModelScope 开源的数据集。The final release of Multi-Crop3D contains 1,658 manually annotated point clouds from 208 plants. Overall statistics are as follows:

cau721582/Multi-Crop3D 是 ModelScope 魔搭社区上的数据集，存储大小 3.3 GB，采用 Apache License 2.0 许可。

- **Repository**: cau721582/Multi-Crop3D
- **License**: Apache License 2.0
- **Storage size**: 3.3 GB
- **Downloads**: 228
- **Stars**: 0
- **Last updated**: 2026-05-07

Source: https://www.modelscope.cn/datasets/cau721582/Multi-Crop3D

---

## Dataset Statistics

The final release of **Multi-Crop3D** contains **1,658** manually annotated point clouds from **208** plants. Overall statistics are as follows:

- Raw point count per sample ranges from **4,112** to **1,152,621**.
- The mean raw point count is **111,955.39**, and the median is **68,986.5**.
- The number of leaf instances per sample ranges from **2** to **42**.
- The mean number of leaf instances is **9.35**, and the median is **7**.

These statistics indicate substantial variation in geometric scale and organ complexity across crops and time points.

### 1. Overall Scale

Multi-Crop3D was constructed through repeated multi-day acquisition of the same plants. Some nominal acquisition time points were removed after 3D reconstruction due to quality issues, such as:

- Incomplete point clouds or insufficient view coverage
- Structural deformation caused by wind disturbance or plant motion
- Heavy background contamination
- Insufficient point density for reliable organ-level annotation

To ensure high-quality organ-level annotation, all samples were subjected to a unified quality control (QC) process before annotation and benchmark construction. Only samples satisfying the predefined QC criteria were retained in the final release.

| Item | Value |
|---|---:|
| Number of crop categories | 4 |
| Number of varieties / groups | 12 |
| Final number of retained plants | 208 |
| Final number of retained point cloud samples | 1,658 |
| Nominal number of acquired samples before QC | 1,760 |
| Number of samples removed during QC | 102 |

### 2. Crop-Level Statistics

| Crop | Raw Point Clouds | Plants | Min Points | Max Points | Mean Points | Median Points | Min Leaf Instances | Max Leaf Instances | Mean Leaf Instances | Median Leaf Instances |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| Soybean | 299 | 30 | 12,148 | 568,337 | 159,825.85 | 113,359 | 2 | 17 | 8.63 | 9 |
| Rice | 196 | 28 | 7,224 | 201,308 | 66,554.32 | 59,776.5 | 2 | 15 | 7.63 | 7 |
| Maize | 857 | 80 | 4,112 | 1,152,621 | 117,391.71 | 54,787 | 3 | 13 | 6.49 | 6 |
| Tomato | 306 | 70 | 16,136 | 205,464 | 79,035.21 | 68,826 | 9 | 42 | 19.15 | 18 |

### 3. Variety / Group-Level Statistics

| Crop | Variety / Group | Final Plants | Final Point Clouds |
|---|---|---:|---:|
| Soybean | D21020 | 10 | 100 |
| Soybean | D21116 | 10 | 100 |
| Soybean | D38 | 10 | 99 |
| Rice | M107 | 10 | 70 |
| Rice | M55 | 9 | 63 |
| Rice | M56 | 9 | 63 |
| Maize | A619 | 20 | 197 |
| Maize | B73 | 30 | 366 |
| Maize | W64A | 30 | 294 |
| Tomato | MicroTom | 19 | 19 |
| Tomato | Saopolo | 26 | 98 |
| Tomato | Starlor | 25 | 189 |

---

## Temporal Acquisition Properties

Multi-Crop3D adopts a repeated acquisition strategy for the same plants over multiple days. Therefore, the dataset contains both cross-crop variation and temporal structural changes.

| Crop | Observation Object | Observation Window | Typical Acquisition Frequency | Temporal Characteristics |
|---|---|---|---|---|
| Maize | Repeated acquisition of the same plants | July 19-August 15 | Approximately daily | Covers 12-15 temporal observations |
| Soybean | Repeated acquisition of the same plants | July 24-August 2 | Approximately daily | Covers about 10 temporal observations |
| Rice | Repeated acquisition of the same plants | July 24-July 30 | Approximately daily | Covers about 7 temporal observations |
| Tomato | Repeated acquisition of the same plants | 25-40 days after emergence | Every 1-2 days | Longitudinal observation over a longer period |

**Notes:**

- Due to QC filtering, the final number of retained time points varies across plants.
- For tomato, orientation markers were used during acquisition to improve consistency across repeated scans.

---

## Official Train/Test Split

### 1. Split Principle

To avoid **temporal leakage**, where different time points of the same plant appear in both training and test sets, Multi-Crop3D adopts a **plant-level split**:

- All time points belonging to the same `plant_id` are assigned exclusively to either the **train** or **test** set.
- The split is designed to maintain coverage across different crops and varieties/groups.
- Since the number of retained time points varies among plants after QC, the file-level counts are not strictly balanced, while the plant-level split protocol remains consistent.

### 2. Plant-Level Split by Variety / Group

| Crop | Variety / Group | Train Plants | Test Plants |
|---|---|---:|---:|
| Soybean | D21020 | 8 | 2 |
| Soybean | D21116 | 8 | 2 |
| Soybean | D38 | 8 | 2 |
| Rice | M107 | 8 | 2 |
| Rice | M55 | 7 | 2 |
| Rice | M56 | 7 | 2 |
| Maize | A619 | 17 | 3 |
| Maize | B73 | 25 | 5 |
| Maize | W64A | 25 | 5 |
| Tomato | MicroTom | 15 | 4 |
| Tomato | Saopolo | 21 | 5 |
| Tomato | Starlor | 20 | 5 |

### 3. Official Split Summary by Crop

| Crop | Train Plants | Test Plants | Train Files | Test Files | Total Files |
|---|---:|---:|---:|---:|---:|
| Soybean | 24 | 6 | 239 | 60 | 299 |
| Rice | 22 | 6 | 154 | 42 | 196 |
| Maize | 67 | 13 | 649 | 208 | 857 |
| Tomato | 56 | 14 | 237 | 69 | 306 |
| **Total** | **169** | **39** | **1,279** | **379** | **1,658** |

---

## Data Acquisition and Crop Growth Conditions

The point clouds were acquired using the MVS64 multi-view imaging system. After reconstruction, HSV-based color thresholding, height-based filtering, and statistical outlier removal were applied to remove most background interference and reconstruction artifacts. The cleaned plant point clouds were then manually annotated in CloudCompare at the organ-instance level and further reviewed for annotation quality.

### Maize / Soybean / Rice Growth Conditions

- Maize and soybean were grown in **20 L white plastic pots**.
- Rice was grown in **7 L black pots**.
- Plants were grown under outdoor natural conditions.
- Soil mixture: **garden soil : peat soil : vermiculite = 5 : 3 : 2**.
- Three seeds were sown per pot, and thinning was performed three days after emergence, retaining one uniformly growing plant per pot.

### Tomato Growth Conditions

- Tomato plants were cultivated on a three-layer greenhouse rack.
- Five grow lights were installed above each layer to control light intensity.
- Temperature range: **18°C-28°C**.
- Light source: **120 W full-spectrum 301B LED**.
- Plants were transplanted into **10.5 cm diameter** pots at the two-leaf stage.
- Substrate mixture: **peat : vermiculite : perlite = 5 : 4 : 1**.
- To maintain consistent orientation across repeated acquisitions, a **2.5 x 2.5 cm yellow tape marker** was attached 3 cm below the pot rim, and the marker was aligned with camera No. 7 during imaging.

---

## Imaging System

All data were acquired using the **MVS64 multi-view stereo reconstruction system**.

### System Components

- 64 cameras (EOS 1300D DSLR)
- 8 terminal controllers
- 4 computing nodes
- 1 master control computer

### Imaging Characteristics

- Synchronized imaging from 64 views
- Over 75% overlap between adjacent views
- Suitable for high-quality plant 3D reconstruction and point cloud acquisition

---

## File Organization

```text
Multi-Crop3D/
├── maize/
│   ├── all-maize.zip          # Raw annotated point clouds
│   └── maize.zip              # Benchmark version, 4,096 points per sample
├── soybean/
│   ├── all-soybean.zip        # Raw annotated point clouds
│   └── soybean.zip            # Benchmark version, 4,096 points per sample
├── rice/
│   ├── all-rice.zip           # Raw annotated point clouds
│   └── rice.zip               # Benchmark version, 4,096 points per sample
├── tomato/
│   ├── all-tomato.zip         # Raw annotated point clouds
│   └── tomato.zip             # Benchmark version, 4,096 points per sample
└── README.md
```

---

## Data Format

The dataset contains two file types:

1. **Raw annotated point clouds**: These files preserve more complete fields and are suitable for customized preprocessing, instance-level analysis, and extended research tasks.
2. **Downsampled point clouds (benchmark version)**: Each sample is uniformly downsampled to **4,096 points**, uses a unified field format, and follows the **official train/test split**.

### 1. Raw Annotated Point Clouds

The raw point cloud fields vary slightly across crops.

#### 1.1 Soybean (`all-soybean.zip`)

| Column | Field | Description |
|---|---|---|
| 1-3 | `x, y, z` | 3D coordinates |
| 4-6 | `r, g, b` | RGB color values |
| 7-9 | `x_norm, y_norm, z_norm` | Normalized 3D coordinates |
| Last column | `label` | Label information; `0` indicates stem, and non-zero values indicate different leaf instances |

#### 1.2 Tomato (`all-tomato.zip`)

| Column | Field | Description |
|---|---|---|
| 1-3 | `x, y, z` | 3D coordinates |
| 4-6 | `r, g, b` | RGB color values |
| 7 | `label` | Label information; `0` indicates stem, and labels such as `1.1` indicate branch-structured leaf identifiers |

#### 1.3 Rice (`all-rice.zip`)

| Column | Field | Description |
|---|---|---|
| 1-3 | `x, y, z` | 3D coordinates |
| 4-6 | `r, g, b` | RGB color values |
| 7 | `label` | Label information; `0` indicates stem, and non-zero values indicate different leaf instances |
| 8-10 | `x_norm, y_norm, z_norm` | Normalized 3D coordinates |

#### 1.4 Maize (`all-maize.zip`)

| Column | Field | Description |
|---|---|---|
| 1-3 | `x, y, z` | 3D coordinates |
| 4-6 | `r, g, b` | RGB color values |
| 7 | `semantic_label` | Semantic label; `1` indicates stem and `2` indicates leaf |
| Last column | `instance_label` | Instance label; `0` indicates stem, and non-zero values indicate different leaf instances |

### 2. Benchmark Version: 4,096 Points per Sample

#### 2.1 Unified Field Format

| Column | Field | Description |
|---|---|---|
| 1-3 | `x, y, z` | 3D coordinates |
| 4-6 | `r, g, b` | RGB color values |
| Last column | `label` | Unified label; `0` indicates stem, and non-zero values indicate different leaf instances |

#### 2.2 Label Unification Rules

To support unified modeling, raw annotations from different crops are mapped into a consistent organ-instance representation:

- **Stem**: uniformly mapped to `0`
- **Leaf instances**: mapped to positive integer IDs
- For maize, `semantic_label` and `instance_label` are jointly used to generate unified instance labels.
- For tomato, branch-structured labels are converted into unified instance IDs in the benchmark version.

#### 2.3 Benchmark Split Organization

```text
train/
test/
```
---
## Citation

If you use this dataset in your research, please cite the corresponding Multi-Crop3D paper and dataset page after the official release.

The following related publication may also be cited if it is relevant to your use of the dataset:

Zhou, J., Zhang, Y., Zhang, M., Zhang, M.Q., Song, Q., Zhu, X., Wang, M. Leveraging time-series point clouds for dynamic crop canopy monitoring: Quantifying phenotypic variability and assessing leaf-level photosynthetic contributions. *Plant Phenomics*, 8, 100194 (2026). https://doi.org/10.1016/j.plaphe.2026.100194

## Contact

For questions about the dataset, please contact the corresponding author listed in the associated paper or the dataset maintainers on the hosting platform.