# Baseline — Cotton / Palmer demo sample

100 synthetic RGB images · 640 × 640 · 80 labeled Palmer plants · 20 negative images

Explore the interactive preview at https://baselineperception.com/sample.

This web download uses JPEG RGB images (quality 92, 4:4:4) at the original
640 × 640 resolution. Labels and lossless PNG masks are unchanged.

## What this demonstrates

Procedurally rendered Palmer amaranth in cotton context, accompanied by detection
boxes and visible plant masks. The labels are generated annotations, not model
predictions. Cotton is intentionally unlabeled in this sample. A negative image
has no labeled Palmer target; it may still contain cotton.

These images already existed in Baseline's September 14, 2026 accepted export.
This package selects 20 complete layouts, evenly spaced through the 50-layout
source, retaining all five views per layout. It is a curated subset, not a fresh
render or an independent dataset. Selection did not use detector scores.

The sample includes first-true-leaf, juvenile, and early-vegetative Palmer assets,
layout/rotation changes, and varied target size and visibility. It uses a fixed
overhead perspective camera and a fixed clear-noon lighting profile over red mud.
It does not demonstrate broad camera, weather, soil, or lighting coverage.

## Files and labels

- `images/`: JPEG RGB images converted from the source PNGs without resizing.
- `labels/`: matching detection text files: `class cx cy width height`, with
  coordinates normalized by image dimensions. Class 0 is Palmer. Negative files
  are empty. This common YOLO data format requires no Ultralytics software.
- `annotations/instances.json`: COCO boxes and uncompressed column-major RLE
  visible instance masks. Category ID 1 is Palmer. Image paths are relative to
  the package root. Original image and annotation IDs are preserved and sparse.
- `masks/`: semantic PNGs (0 background including cotton, 1 Palmer) and 16-bit
  instance PNGs. Nonzero instance IDs match COCO `instance_mask_value`; do not
  interpret them as class IDs or convert these masks to 8-bit.
- `metadata/`: per-image camera intrinsics, camera transform, lighting settings,
  layout/view groups, label thresholds, and source hashes.
- `manifest.json`, `validation.json`, `SHA256SUMS.txt`: selection and integrity records.

Boxes use a lower coverage threshold than masks (0.1 versus 0.5), so a box may
extend slightly beyond the visible mask. Masks cover visible plant pixels, not
hidden plant portions. There are no separate leaf-instance or disease labels in
this package. Small-target and partial-view cases are included.

## Try the files

Run `python3 inspect_sample.py` from any directory. It uses only Python's standard
library, checks the package hashes and basic COCO relationships, and prints the
sample counts. To load the COCO JSON yourself:

```python
import json
from pathlib import Path
root = Path("Baseline_Cotton_Palmer_Sample_v1")
coco = json.loads((root / "annotations/instances.json").read_text())
image_path = root / coco["images"][0]["file_name"]
```

## Evaluation scope

This is an inspection and integration sample, not a benchmark or a ready-made
train/validation/test experiment. Related views share a `layout_group_id`; keep
them together if splitting. All images share the same asset family and capture
profile. This subset overlaps Baseline's existing synthetic training material.
A field-performance claim needs a separate, untouched real-image evaluation set
and a pre-agreed comparison of real-only versus real-plus-synthetic training.
No biological validation, accuracy improvement, or customer camera match is claimed.

## About the sample

Produced by Baseline Perception LLC using Blender and Baseline's procedural plant
library. Ground appearance uses Poly Haven's Red Mud Stones by Rob Tuytel (CC0).
No external field photographs, model weights, or source 3D assets are included.
Free for internal research and evaluation, including within companies.
See LICENSE.txt for the complete permissions. Production use, model distribution
and dataset redistribution require separate permission.
For another crop/camera: adad@baselineperception.com.

https://baselineperception.com/