PlantMetricDepth

PlantMetricDepth is a multimodal plant dataset designed for metric monocular depth estimation (MDE) and related plant analysis tasks.

The dataset provides paired stereo RGB images, disparity maps, generated metric depth maps, and plant segmentation masks collected across 15 acquisition days.

The metric depth maps provide dense per-pixel depth supervision in centimetres, enabling models to learn metric depth from a single RGB image at inference time without requiring a stereo camera.

Dataset Contents

The repository is organised into 15 folders:

PlantMetricDepth/
├── day_1/
├── day_2/
├── day_3/
│   ...
└── day_15/

Each acquisition day contains five modalities:

day_X/
├── Left/
├── Right/
├── Disparity/
├── Depth/
└── Seg_mask/

Modalities

Folder	Description
Left/	Left-view RGB plant images
Right/	Corresponding right-view RGB images
Disparity/	Stereo disparity maps
Depth/	Generated metric depth maps
Seg_mask/	Plant segmentation masks

The RGB images have a spatial resolution of 650 × 650 pixels.

---

Metric Depth Maps

Each sample in the Depth directory contains:

plant1_day1.npy
plant1_day1.png

The two files have different purposes:

• .npy — numerical metric depth map used for training and evaluation.
• .png — visualisation of the corresponding depth map.

The .npy files contain dense floating-point depth values expressed in centimetres.

For quantitative experiments, use the .npy depth maps rather than the PNG visualisations.

Loading a depth map

import numpy as np

depth = np.load(
    "PlantMetricDepth/day_1/Depth/plant1_day1.npy"
)

print(depth.shape)
print(depth.dtype)
print("Minimum depth:", depth.min(), "cm")
print("Maximum depth:", depth.max(), "cm")

---

Metric Depth Generation

Metric depth supervision was generated from the stereo image pairs using FoundationStereo.

The processing pipeline consists of three main stages:

Left + Right RGB
       │
       ▼
FoundationStereo
       │
       ▼
Stereo Disparity
       │
       ▼
Inverse-disparity depth
       │
       ▼
Floor-plane metric scaling
       │
       ▼
Metric Depth (cm)

FoundationStereo was used in its pre-trained configuration without dataset-specific fine-tuning.

Given a disparity map (d), an inverse-disparity representation is obtained as:

$$D_{\mathrm{r}\mathrm{e}\mathrm{l}}=\frac{1}{d+ϵ},D\_\{\backslash mathrm\{rel\}\}\; =\; \backslash frac\{1\}\{d\; +\; \backslash epsilon\},$$

where (D_{\mathrm{rel}}) represents relative depth and (\epsilon) prevents division by zero.

Metric scale is recovered using the known camera-to-floor distance:

$$H=150\mathrm{c}\mathrm{m}\mathrm{.}H\; =\; 150\backslash \; \backslash mathrm\{cm\}.$$

The median relative depth over the floor reference region is calculated as:

$$r_{\mathrm{f}\mathrm{l}\mathrm{o}\mathrm{o}\mathrm{r}}=\mathrm{median}(D_{\mathrm{r}\mathrm{e}\mathrm{l}}(p)),p\in M_{\mathrm{f}\mathrm{l}\mathrm{o}\mathrm{o}\mathrm{r}}\mathrm{.}r\_\{\backslash mathrm\{floor\}\}\; =\; \backslash operatorname\{median\}\; \backslash left(\; D\_\{\backslash mathrm\{rel\}\}(p)\; \backslash right),\; \backslash quad\; p\; \backslash in\; M\_\{\backslash mathrm\{floor\}\}.$$

The metric scale factor is then:

$$s=\frac{H}{r_{\mathrm{f}\mathrm{l}\mathrm{o}\mathrm{o}\mathrm{r}}},s\; =\; \backslash frac\{H\}\{r\_\{\backslash mathrm\{floor\}\}\},$$

and the final metric depth is:

$$D_{\mathrm{m}\mathrm{e}\mathrm{t}\mathrm{r}\mathrm{i}\mathrm{c}}=s{D}_{\mathrm{r}\mathrm{e}\mathrm{l}}\mathrm{.}D\_\{\backslash mathrm\{metric\}\}\; =\; sD\_\{\backslash mathrm\{rel\}\}.$$

The resulting depth maps therefore provide centimetre-scale dense depth supervision.

---

Example Sample Pairing

Files belonging to the same plant and acquisition day share the same sample identifier.

For example:

day_1/
├── Left/
│   └── plant1_day1.png
│
├── Right/
│   └── plant1_day1.png
│
├── Disparity/
│   └── plant1_day1.*
│
├── Depth/
│   ├── plant1_day1.npy
│   └── plant1_day1.png
│
└── Seg_mask/
    └── plant1_day1.*

This naming convention allows the modalities to be paired directly.

---

Downloading the Dataset

The Hugging Face repository ID is:

BashayerAA/PlantMetricDepth

Option 1 — Hugging Face CLI

This is the recommended method for downloading the complete dataset.

1. Install Hugging Face Hub

pip install -U huggingface_hub

2. Download the dataset

hf download BashayerAA/PlantMetricDepth \
    --repo-type dataset \
    --local-dir PlantMetricDepth

The complete repository will be downloaded into:

PlantMetricDepth/

The original directory structure will be preserved.

---

Option 2 — Download with Python

The complete dataset can also be downloaded programmatically.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="BashayerAA/PlantMetricDepth",
    repo_type="dataset",
    local_dir="PlantMetricDepth"
)

After completion:

PlantMetricDepth/
├── day_1/
├── day_2/
├── ...
└── day_15/

---

Download a Single Acquisition Day

If the complete dataset is not required, individual acquisition days can be downloaded.

For example, to download only day_1:

hf download BashayerAA/PlantMetricDepth \
    --repo-type dataset \
    --include "day_1/*" \
    --local-dir PlantMetricDepth

---

Download Only Metric Depth Maps

To download only the numerical metric depth maps:

hf download BashayerAA/PlantMetricDepth \
    --repo-type dataset \
    --include "day_*/Depth/*.npy" \
    --local-dir PlantMetricDepth

---

Download Only Left RGB Images

For monocular RGB input images only:

hf download BashayerAA/PlantMetricDepth \
    --repo-type dataset \
    --include "day_*/Left/*" \
    --local-dir PlantMetricDepth

---

Loading RGB and Depth Pairs

A simple example for loading a left RGB image with its corresponding metric depth map is:

from PIL import Image
import numpy as np

rgb_path = "PlantMetricDepth/day_1/Left/plant1_day1.png"
depth_path = "PlantMetricDepth/day_1/Depth/plant1_day1.npy"

rgb = Image.open(rgb_path).convert("RGB")
depth = np.load(depth_path)

print("RGB size:", rgb.size)
print("Depth shape:", depth.shape)
print("Depth range:", depth.min(), depth.max(), "cm")

For monocular depth estimation, the typical training pair is:

Input  : Left RGB image
Target : Metric depth (.npy)

The right RGB image and disparity map are therefore not required during monocular inference.

---

Intended Uses

PlantMetricDepth can be used for research involving:

• Metric monocular depth estimation
• Plant phenotyping
• Agricultural computer vision
• RGB-to-depth prediction
• Stereo-to-monocular knowledge transfer
• Depth-assisted plant classification
• Multimodal plant analysis
• Evaluation of dense depth-estimation models

---

Dataset Characteristics

Property	Description
Domain	Greenhouse plant imagery
Acquisition period	15 days
Image resolution	650 × 650
RGB views	Left and right
Depth type	Dense metric depth
Depth unit	Centimetres
Stereo information	Disparity maps
Segmentation	Plant segmentation masks
Monocular input	Left RGB
Metric supervision	Generated depth map

---

Important Notes

The metric depth maps are generated supervision rather than direct measurements from an active depth sensor.

Metric scaling relies on the known 150 cm camera-to-floor distance and the floor reference region.

---

Licence

This dataset is released under the Apache License 2.0.

---

Citation

If you use PlantMetricDepth in your research, please cite the associated publication when available.

@dataset{plantmetricdepth,
  author    = {Bashayer Abdallah, Shan E Ahmed Raza},
  title     = {PlantMetricDepth: A Plant Dataset with Generated Metric Depth Supervision},
  publisher = {Hugging Face},
  year      = {2026}
}


---

# Contact

For questions regarding the dataset, methodology, or research use, please use the **Community** section of the PlantMetricDepth Hugging Face repository.