Sungjay/3DKoreanMelon

Korean Melon 3D Growth Sequences

Per-fruit 3D observation sequences of Korean melon (Cucumis melo L. var. makuwa) grown on the plant, each paired with a post-harvest scan of the same fruit and with vernier caliper measurements from visits with available readings.

Fruits were revisited every two to three days over a full growing period and imaged in place, so each sequence follows one identified fruit as it enlarges while foliage occludes a different part of it at each visit.

Fruits	210
Partial point clouds	2,738 (1–16 sessions per fruit, 13 on average)
Sessions with caliper readings	2,312
Post-harvest reference scans	210
Seasons / greenhouses / cultivars	2 / 3 / 2
Size	~334 MB

Layout

processed/<sample_id>/<YYYYMMDD>.ply    partial point cloud, one per imaging session
gt/<sample_id>/pointcloud.ply           post-harvest reference scan of the same fruit
metadata/sessions.csv                   one row per fruit-session, all measurements
metadata/splits.json                    train / validation / test fruit ids
metadata/camera/                        depth and colour intrinsics, depth-to-colour extrinsics
tools/compute_volume.py                 recomputes harvest_volume_ml from a reference scan

Prefix	Greenhouse	Imaging window	Cultivar	Fruits	Split
A1–A61	A	2025, 1–29 May	Eliteggul	61	train, validation
B1–B69	B	2026, 11 Mar – 16 Apr	Alchanggul	69	train, validation
C1–C80	C	2026, 11 Mar – 16 Apr	Alchanggul	80	test

Greenhouse C was held out entirely for testing.

Point clouds

processed/ — one cloud per fruit per imaging session, in metres, carrying XYZ and RGB. Each is the masked fruit region of a single RGB-D frame, back-projected with the factory intrinsics and translated so its centroid sits at the origin. These are partial views: foliage, the trellis, and the fruit's own far side are all missing, and how much is missing varies session to session.

gt/ — one reference scan per fruit, in metres, XYZ only, recentred and aligned to a canonical orientation. Each was reconstructed from 108 images of the detached fruit on a turntable.

metadata/sessions.csv

One row per fruit-session. A session appears whether or not it produced a point cloud, so check has_pointcloud and has_caliper before reading the corresponding columns.

Column	Meaning
sample_id	public identifier, e.g. C12
source_id	identifier used in the accompanying code
greenhouse, season, cultivar, split	per-fruit attributes, repeated on each row
session_date	imaging date, YYYYMMDD
is_last_session	1 on the final session, the one nearest harvest
has_pointcloud	1 if processed/<sample_id>/<session_date>.ply exists
n_points	points in that cloud
passes_quality_filter	1 if n_points >= 300, the threshold used in the accompanying study
has_caliper	1 if the fruit was measured at this session
height_mm, width_mm	vernier caliper readings
ellipsoid_ml	prolate-ellipsoid volume from those two readings
corrected_gt_ml	ellipsoid_ml scaled by alpha_correction, a scaled caliper size reference for sessions before harvest
harvest_volume_ml	volume of the post-harvest scan, as the convex hull of gt/<sample_id>/pointcloud.ply
harvest_mass_g	mass at harvest, where recorded
alpha_correction	harvest_volume_ml divided by the last available caliper ellipsoid_ml
n_sessions	sessions of this fruit that produced a point cloud

The final session uses the reconstructed harvest volume. Earlier references scale the measured caliper ellipsoid volumes; they are not temporal interpolations. This retrospective evaluation reference uses harvest information and is not an inference input. Any harvest volume can be checked against the cloud it came from with python tools/compute_volume.py --all.

Loading

import open3d as o3d
import pandas as pd

sessions = pd.read_csv("metadata/sessions.csv", dtype={"session_date": str})

fruit = sessions[(sessions.sample_id == "C12") & (sessions.has_pointcloud == 1)]
for date in fruit.sort_values("session_date").session_date:
    pcd = o3d.io.read_point_cloud(f"processed/C12/{date}.ply")

reference = o3d.io.read_point_cloud("gt/C12/pointcloud.ply")

Collection

Grown under vertical downward training at the Seongju Korean Melon and Vegetable Research Institute, Gyeongsangbuk-do Agricultural Research and Extension Services, Republic of Korea. An Orbbec Femto Mega time-of-flight sensor on a rail-mounted platform imaged the aisles from 500–800 mm. RGB-D frame saving was synchronized with detection of floor-mounted ArUco markers; fruits were individually labeled across visits. SAM 2 masks used point prompts with inspection and correction during the first five manual-labeling visits, followed by automatic tracking. Depth outside 400–900 mm was discarded before back-projection. The last on-plant observation was immediately before harvest; the in-lab reference images were captured within six hours of harvest. Reference reconstruction used COLMAP, scaled through ArUco corner triangulation.

Both seasons use metres. Acquisition used WFOV depth (1024 × 1024) and 1920 × 1080 RGB in 2025, and NFOV depth (640 × 576) and 1280 × 720 RGB in 2026. Greenhouse A reference scans were resampled to 20,000 points, while B and C retain their reconstructed resolution. 28 fruits have no caliper readings.

Raw RGB frames, depth maps, and masks are not included; the point clouds are the processed form the accompanying study consumed. Trained weights are released with the code.

Paper evaluation supplement

metadata/paper_evaluation.csv adds the exact paper evaluation cohort and its original screened-Poisson volume references, alongside the existing convex-hull references in sessions.csv. There are 79 test fruits (79 final and 976 intermediate observations) and 18 validation fruits (18 final and 182 intermediate observations). Among all released clouds, 2,513 meet the 300-point threshold; evaluation additionally requires available volume references and at least three eligible visits. Intermediate visits require caliper readings; final visits use harvest geometry and may lack a same-day caliper reading.

benchmarks/score_reference/ contains archived predictions, fixed occlusion masks, and analysis tables for the accompanying code's portable reanalysis scripts. Original point clouds, splits, and sessions.csv are preserved. The bundle distinguishes archived-paper reanalysis from new inference on the public point clouds.

Citation

The accompanying paper is not published yet. A preprint reference will be added here once it is posted, and replaced by the journal reference after that; until then, please cite this dataset by its DOI, 10.57967/hf/9982.

@unpublished{kim2026score,
  title  = {Temporal latent fusion for sequential 3D shape completion in
            on-plant Korean melon growth monitoring},
  author = {Kim, Sungjay and Blok, Pieter M. and Xin, Xianghui and Kim, Gyumin and
            Go, Yeongjun and Ryu, Jiwon and Kim, Sang-Yeon and Lee, Chang-Hyup and
            Kim, Ghiseok},
  year   = {2026},
  note   = {Manuscript in preparation}
}

Code and trained weights: https://github.com/sungjay-kim/SCoRe

License

CC BY 4.0. Use it for anything, including commercially, with attribution.