PEAR

Pose and dEformation of Agricultural pRoduce

Overview

PEAR is an RGB-D benchmark designed to study the instance-specific geometric deformation of agricultural produce, while its object-pose annotations also allow it to be used independently for conventional 6D pose estimation and evaluation.

PEAR accompanies Mind the Shape Gap: A Benchmark and Baseline for Deformation-Aware 6D Pose Estimation of Agricultural Produce (IROS 2026). The paper also introduces SEED (Simultaneous Estimation of posE and Deformation), an RGB-only method that jointly estimates 6D pose and explicit lattice deformation from a category template.

Dataset at a Glance

Property	Value
Produce categories	8
Categories	Apple, avocado, banana, carrot, lemon, long pepper, pear, pumpkin
Real produce instances	39
Real annotated scenes	117
Valid real RGB-D frames	9,607
Capture conditions per instance	3
Real image resolution	1280 x 960
Pose annotations	Object-to-camera 6D pose per valid frame
Geometry	Scanned mesh for every real instance

For each real produce instance, the scene suffix describes the capture condition:

• _0: isolated produce without added clutter;
• _1: produce in a cluttered scene;
• _2: produce with occlusion.

Only frames listed in each scene's valid_frame_ids are part of the annotated dataset. Frames rejected during annotation and quality control are not included.

Dataset Structure

PEAR/
├── README.md
├── dataset_manifest.json
├── frames.csv
├── scenes/
│   └── scene_XXX_Y/
│       ├── scene_manifest.json
│       ├── cam_K.txt
│       ├── object_pose_m2w.txt
│       ├── rgb/
│       │   └── frame_XXXXX.png
│       ├── depth/
│       │   └── frame_XXXXX.png
│       ├── masks/
│       │   └── frame_XXXXX.png
│       ├── poses/
│       │   └── frame_XXXXX.txt
│       ├── camera_extrinsics/
│       │   └── frame_XXXXX.txt
│       └── mesh/
│           ├── instance.obj
│           ├── instance.mtl
│           └── texture.png
├── pointclouds/
│   ├── scene_XXX_Y.ply
│   └── scene_XXX_Y.json
└── occl_index/
    ├── all_occluded_frames.csv
    └── scene_XXX_2.csv

dataset_manifest.json indexes all included scenes. frames.csv provides one row per valid frame, including its scene, original frame ID, modality paths, occlusion status, and annotation paths. Frame IDs are preserved from the source recordings and are not renumbered.

Annotations

RGB

rgb/frame_XXXXX.png contains the original color image.

Corrected depth

depth/frame_XXXXX.png is a single-channel uint16 PNG. Values are integer millimetres, and zero denotes invalid or unavailable depth.

Object masks

masks/frame_XXXXX.png is a single-channel binary PNG:

• 0: background;
• 255: object.

Each exported frame contains one canonical object mask.

Object-to-camera poses

poses/frame_XXXXX.txt stores T_m2c, the 4 x 4 homogeneous transform from the original mesh coordinate system to the corresponding camera coordinate system. Translations are in metres.

For a column-vector mesh point p_mesh:

p_camera = T_m2c @ p_mesh

Camera extrinsics

camera_extrinsics/frame_XXXXX.txt stores T_c2w, the final corrected camera-to-world transform. It maps camera coordinates into the shared, gauge-rectified scene world frame:

p_world = T_c2w @ p_camera

Shared object pose

object_pose_m2w.txt stores T_m2w, the shared mesh-to-world pose for the scene. The exported transforms satisfy:

T_m2c = inverse(T_c2w) @ T_m2w
T_m2w = T_c2w @ T_m2c

Camera intrinsics

cam_K.txt stores the 3 x 3 pinhole intrinsic matrix:

fx  0 cx
 0 fy cy
 0  0  1

Instance geometry

The mesh/ directory contains one OBJ mesh and its required material and texture dependencies. Mesh vertices are stored in millimetres. Convert them to metres before applying an annotation:

mesh_vertices_m = mesh_vertices_mm * 0.001

Point clouds

pointclouds/scene_XXX_Y.ply is a binary little-endian PLY containing the object-depth observations from every valid frame in the scene. Point clouds are generated by:

1. eroding the binary object mask by 10 pixels;
2. retaining corrected depths between 50 mm and 1,200 mm;
3. back-projecting depth with cam_K.txt;
4. transforming points with the final T_c2w;
5. attaching the corresponding RGB value and source frame ID.

Occlusion indexes

occl_index/scene_XXX_2.csv lists valid frames manually marked as occluded for that scene. occl_index/all_occluded_frames.csv combines these records and adds the scene identifier.

Coordinate Conventions

PEAR uses the OpenCV camera convention:

• +x: image right;
• +y: image down;
• +z: forward from the camera.

Matrices are plain-text, row-major 4 x 4 homogeneous matrices and use the column-vector mathematical convention:

p_destination = T_destination_from_source @ [x, y, z, 1]^T

For points stored as rows in NumPy:

points_destination = points_source @ T[:3, :3].T + T[:3, 3]

Units are:

• pose and extrinsic translations: metres;
• point-cloud coordinates: metres;
• raw OBJ vertices: millimetres;
• depth PNG values: integer millimetres;
• rotation matrices: dimensionless.

Download

PEAR is distributed independently of the SEED implementation through the Hugging Face dataset repository:

https://huggingface.co/datasets/enci2/PEAR

The complete dataset can be downloaded without cloning or installing SEED. While the repository is private, authenticate with an account that has been granted access. Authentication will not be required after the dataset is made public.

Command-line download

python -m pip install -U huggingface_hub
hf auth login
hf download enci2/PEAR --repo-type dataset --local-dir PEAR

To download one scene instead of the complete dataset:

hf download enci2/PEAR \
  --repo-type dataset \
  --include "README.md" \
  --include "dataset_manifest.json" \
  --include "frames.csv" \
  --include "scenes/scene_003_0/**" \
  --local-dir PEAR_scene_003_0

Python download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="enci2/PEAR",
    repo_type="dataset",
    local_dir="PEAR",
)

Usage

The files can be consumed directly using dataset_manifest.json, frames.csv, and the per-scene manifests. The optional loader and validation utilities are maintained separately in the PEAR directory of the project repository:

git clone https://github.com/kvantonikolas/PEAR-SEED.git
cd PEAR-SEED/PEAR
python -m pip install -e .

Installing this loader does not download SEED model weights and is not required to access the dataset. The following standalone example uses common Python packages to load one frame and project its instance mesh with the ground-truth pose:

from pathlib import Path

import cv2
import numpy as np
import trimesh

scene = Path("PEAR/scenes/scene_003_0")
frame_id = 0
name = f"frame_{frame_id:05d}"

rgb = cv2.cvtColor(
    cv2.imread(str(scene / "rgb" / f"{name}.png")),
    cv2.COLOR_BGR2RGB,
)
depth_mm = cv2.imread(
    str(scene / "depth" / f"{name}.png"), cv2.IMREAD_UNCHANGED
)
mask = cv2.imread(
    str(scene / "masks" / f"{name}.png"), cv2.IMREAD_GRAYSCALE
) > 0

K = np.loadtxt(scene / "cam_K.txt").reshape(3, 3)
T_m2c = np.loadtxt(scene / "poses" / f"{name}.txt").reshape(4, 4)
T_c2w = np.loadtxt(
    scene / "camera_extrinsics" / f"{name}.txt"
).reshape(4, 4)
T_m2w = np.loadtxt(scene / "object_pose_m2w.txt").reshape(4, 4)

mesh_path = next((scene / "mesh").glob("*.obj"))
mesh = trimesh.load(mesh_path, force="mesh", process=False)
vertices_m = np.asarray(mesh.vertices, dtype=np.float64) * 0.001
vertices_camera = vertices_m @ T_m2c[:3, :3].T + T_m2c[:3, 3]

visible = vertices_camera[:, 2] > 0
xyz = vertices_camera[visible]
uv = np.column_stack((
    K[0, 0] * xyz[:, 0] / xyz[:, 2] + K[0, 2],
    K[1, 1] * xyz[:, 1] / xyz[:, 2] + K[1, 2],
))

assert np.allclose(np.linalg.inv(T_c2w) @ T_m2w, T_m2c, atol=1e-8)

Consult dataset_manifest.json, frames.csv, and each scene_manifest.json instead of assuming that every numerical frame ID exists.

License

The PEAR dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You may share and adapt the dataset, including for commercial use, provided that appropriate credit is given, a link to the license is provided, and any changes are indicated. Please provide attribution by citing the PEAR paper below. See LICENSE for the complete legal terms.

The dataset license applies to the released data and annotations. Software in the PEAR-SEED GitHub repository is distributed separately under the Apache License 2.0.

Citation

If you use PEAR, please cite the accompanying paper:

@inproceedings{Chatzis2026PEAR,
  title={Mind the Shape Gap: A Benchmark and Baseline for Deformation-Aware 6D Pose Estimation of Agricultural Produce},
  author={Nikolas Chatzis and Angeliki Tsinouka and Katerina Papadimitriou and Niki Efthymiou and Marios Glytsos and George Retsinas and Paris Oikonomou and Gerasimos Potamianos and Petros Maragos and Panagiotis Paraskevas Filntisis},
  booktitle={Proceedings of the International Conference on Intelligent Robots and Systems (IROS)},
  year={2026}
}

Links

• Paper: https://arxiv.org/abs/2603.27429
• GitHub: https://github.com/kvantonikolas/PEAR-SEED
• Project page: https://kvantonikolas.github.io/PEAR-SEED/
• Hugging Face dataset: https://huggingface.co/datasets/enci2/PEAR