field2sim: GreenSplat capsicum release (v1.0.0)
3D Gaussian Splatting reconstructions of five greenhouse capsicum (sweet pepper) group captures from 2026-05-05. Each scene holds only the plant gaussians, with per-gaussian semantic and fruit-instance annotations. All units are metres.
| scene | gaussians | fruits | cameras | m / SfM unit | ply GB |
|---|---|---|---|---|---|
| capsicum-20260505-orange-group-s01 | 5,666,796 | 419 | 1944 | 0.6759 | 1.43 |
| capsicum-20260505-orange-group-s02 | 5,645,588 | 690 | 3681 | 0.6590 | 1.42 |
| capsicum-20260505-red-group-s01 | 4,100,722 | 199 | 4087 | 0.6127 | 1.03 |
| capsicum-20260505-red-group-s02 | 3,371,665 | 76 | 1548 | 0.3904 | 0.85 |
| capsicum-20260505-yellow-group-s01 | 3,132,605 | 265 | 875 | 0.3423 | 0.79 |
Download
hf download dtabaa/field2sim --repo-type dataset --local-dir field2sim
cd field2sim && sha256sum -c SHA256SUMS
python devkit/example_counting.py capsicum-20260505-red-group-s02
Layout
<scene>/
gaussians.ply 3DGS gaussians + annotation columns (metres)
cameras.json registered cameras, nerfstudio transforms.json schema
metadata.json counts, scale, extents, per-fruit boxes
devkit/ field2sim.py (loader + renderer), example_counting.py
scenes.json index of the scenes
SHA256SUMS sha256sum -c SHA256SUMS
gaussians.ply
A binary little-endian PLY with a single vertex element. The standard 3DGS
columns come first, in the INRIA order: x y z, f_dc_0..2,
f_rest_0..44, opacity, scale_0..2, rot_0..3. They are stored before
activation: logit opacity, log scales, an unnormalised wxyz quaternion, and
SH degree 3 with the SH coefficients channel-major. The annotations follow as
extra columns:
| column | type | meaning |
|---|---|---|
semantic |
int32 | 0 = plant (non-fruit), 1 = fruit |
instance |
int32 | manually corrected fruit id 0..n_fruits-1; -1 = not fruit |
fruit_score |
float32 | SAM3 "pepper" masks uplifted to 3D (LUDVIG), in [0, 1] |
pred_instance |
int32 | automatic instance clustering before correction; -1 = none |
3DGS parsers select columns by name, so they ignore the extra columns. Every
column is 4 bytes wide, and no extra column name starts with rot, scale_ or
f_, the prefixes the reference loader matches on. The release was checked to
load in the INRIA reference GaussianModel.load_ply, in plyfile, in gsplat
and in Spark. The header comment lines repeat this table.
cameras.json
This file follows the nerfstudio transforms.json schema, with per-frame
PINHOLE intrinsics (fl_x fl_y cx cy w h). transform_matrix is the
camera-to-world matrix in OpenGL camera axes (+x right, +y up, +z back). For
OpenCV axes, right-multiply it by diag(1,-1,-1,1).
The world frame is the frame of gaussians.ply: the COLMAP SfM frame scaled to
metres. The intrinsics refer to the COLMAP-undistorted images at 1600 px on the
long side. The images themselves are not part of this release; file_path
gives their names.
Each frame also has exposure, the 3x4 affine colour correction learnt per
image in training. To compare a render with its photograph, use
rgb_photo ≈ rgb_render @ E[:, :3] + E[:, 3], with rgb as a row vector in [0, 1].
metadata.json
counts: numbers of gaussians, fruits, fruit gaussians and cameras.units:metres_per_sfm_unitand how it was measured. Known 0.15 m reference squares were measured in the SfM cloud, giving 12–16 sides and a relative SD of 1.8–3.2 %.orientation.estimated_gravity_down: the mean image-down axis of the cameras, about +y. The frame is not gravity-aligned.extents: AABBs of all gaussian centres, of the 1st–99th percentiles, of the fruits and of the camera centres.fruits[]: one entry per fruit withid(theinstancevalue),source_label,n_gaussians,retained_fraction,centroid,aabb, andobb {center, rotation, extent}. The OBBrotationcolumns are the box axes in world space, sorted by decreasingextent(full side lengths).
Devkit: field2sim
devkit/field2sim.py is a single-file API (numpy, torch, gsplat) that loads a
scene and renders RGB, depth, semantic and fruit-instance images from any camera:
import sys; sys.path.insert(0, "devkit")
from field2sim import Scene
scene = Scene("capsicum-20260505-red-group-s02")
for view in scene.render_sequence(c.scaled(0.5) for c in scene.cameras[::25]):
print(view.camera.name, view.visible_instances(min_pixels=50))
devkit/example_counting.py builds ground truth for counting over a sequence
(fruits visible per frame, distinct fruits seen so far). It uses two sequences:
the original capture trajectory and a synthetic pass along the row.
Known caveats
- The boxes enclose the reconstructed fruit surface, mostly the side the cameras saw, so they underestimate the depth of a fruit.
- Restricting the scene to plants removed some annotated fruits that lay
outside the plant region: 19 in orange-s01, 1 in orange-s02, 50 in
red-s01 and 3 in red-s02. In orange-s01, three fruits were partly removed;
their
retained_fractionis below 1. - 19 fruits have fewer than 50 gaussians. They are probably fragments; filter
them on
n_gaussiansif needed. fruit_scoreandpred_instanceare model outputs.semanticandinstanceare the manually corrected ground truth.
- Downloads last month
- 69