多作物3D点云数据集

## 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.