SawitMVC

SawitMVC is a multi-view oil palm fruit bunch detection and counting dataset. It contains expert-reviewed YOLO annotations and per-tree JSON ground truth for counting unique fruit bunches across 4-8 camera views.

Dataset Summary

Property	Value
Trees	953 (DAMIMAS: 854, LONSUM: 99)
Images	3,992 (960 x 1280 px, JPEG)
Views per tree	4 sides (45 trees have 8 sides)
Annotation format	YOLO v8 labels + JSON ground truth
Classes	4 maturity levels (B1-B4)
Unique bunches (GT)	9,823

Split

Split	Trees	Images
train	716	3,000
val	96	404
test	141	588

Split assigned at tree level — all sides (4 or 8) of a tree go to the same split. Stratified by variety × dominant maturity class × view-count (seed=42, ratio 75/10/15).

Dataset Statistics

Class Distribution (unique bunches)

Class	Count	%	Description
B1	954	9.7%	Ripe — least frequent
B2	1,791	18.2%	Transitioning
B3	5,067	51.6%	Unripe — dominant class
B4	2,011	20.5%	Very unripe
Total	9,823	100%	Unique bunches (GT)

Raw detections across all views: 18,540 → duplication ratio 1.887× (same bunch seen from multiple sides).

Per-Split Class Distribution

Split	Trees	B1	B2	B3	B4	Total bunches
train	716	739	1,337	3,818	1,533	7,427
val	96	95	198	504	195	992
test	141	120	256	745	283	1,404

Bunches per Tree

Metric	Value
Min	0
Median	10
Mean	10.31
Max	22
Std dev	3.74

Bunch Appearance Distribution

How many camera sides each unique bunch is visible from:

Visible from N sides	Bunches	%
1	2,495	25.4%
2	6,264	63.8%
3	834	8.5%
4	147	1.5%
5	71	0.7%
6	12	0.1%

Most bunches (63.8%) appear in exactly 2 adjacent sides, consistent with the camera capture geometry.

Tasks

1. Object detection: detect and classify oil palm fruit bunches in each image.
2. Multi-view counting: use JSON ground truth to count each physical bunch once even when it appears in multiple camera views.

Maturity Classes

Class ID	Label	Stage	Description
0	B1	Ripe	Red, large, round; optimal harvest stage
1	B2	Transitioning	Dark fruit transitioning to red
2	B3	Unripe	Black, spiny, elongated
3	B4	Very unripe	Small, deeply positioned, black to green

Biological order: B1 -> B2 -> B3 -> B4 from most ripe to least ripe.

Sample Visualization

Each color represents one unique bunch. The same color across panels means the same physical bunch appears from multiple sides.

4-view tree:

8-view tree:

Dataset Structure

SawitMVC/
|-- images/                    # 3,992 images, flat structure
|-- labels/                    # 3,992 YOLO .txt files, flat structure
|-- json/                      # 953 JSON ground-truth files, one per tree
|-- data/
|   `-- ground_truth.parquet   # Per-tree ground-truth summary
|-- data.yaml                  # YOLO dataset config
|-- split_manifest.csv         # Tree-level split and stratification metadata
`-- croissant.json             # ML Croissant metadata

File Naming

DAMIMAS_A21B_0001_1.jpg  -> variety=DAMIMAS, code=A21B, tree=0001, side=1
DAMIMAS_A21B_0001_1.txt  -> YOLO label for the same image
DAMIMAS_A21B_0001.json   -> ground truth for all views of tree 0001

YOLO Label Format

# class_id  cx_norm  cy_norm  w_norm  h_norm
2           0.660417 0.408203 0.056250 0.041406
1           0.622396 0.443750 0.098958 0.087500

Coordinates are normalized to [0, 1] relative to a 960 x 1280 image. Class IDs are 0=B1, 1=B2, 2=B3, 3=B4.

JSON Ground Truth Format

{
  "version": 4,
  "tree_id": "DAMIMAS_A21B_0001",
  "split": "train",
  "metadata": {
    "date": "2026-05-16",
    "variety": "DAMIMAS"
  },
  "images": {
    "side_1": {
      "filename": "DAMIMAS_A21B_0001_1.jpg",
      "side_index": 0,
      "side_label": "Side 1",
      "bbox_count": 5,
      "annotations": [
        {
          "box_index": 0,
          "class_id": 2,
          "class_name": "B3",
          "bbox_yolo": [0.660417, 0.408203, 0.05625, 0.041406]
        }
      ]
    }
  },
  "bunches": [
    {
      "bunch_id": 1,
      "class": "B3",
      "appearance_count": 2,
      "appearances": [
        {"side": "side_1", "side_index": 0, "box_index": 0},
        {"side": "side_2", "side_index": 1, "box_index": 2}
      ]
    }
  ],
  "summary": {
    "total_unique_bunches": 8,
    "total_detections": 17,
    "duplicates_linked": 9,
    "by_class": {"B1": 1, "B2": 2, "B3": 5, "B4": 0},
    "by_side": {"side_1": 5, "side_2": 4, "side_3": 4, "side_4": 4}
  }
}

summary.by_class is the ground truth for counting evaluation. _confirmedLinks stores annotator-confirmed cross-view links using numeric sideA, sideB, bboxIdA, and bboxIdB references.

Parquet Ground Truth

data/ground_truth.parquet contains one row per tree.

Columns: tree_id, split, variety, num_sides, total_unique_bunches, B1, B2, B3, B4, total_detections, duplicates_linked

Example query:

SELECT variety, AVG(total_unique_bunches) AS avg_bunches,
       SUM(B1) AS total_B1, SUM(B2) AS total_B2,
       SUM(B3) AS total_B3, SUM(B4) AS total_B4
FROM ground_truth
GROUP BY variety;

Usage

from datasets import load_dataset

ds = load_dataset("ULM-DS-Lab/SawitMVC", data_dir="images")
gt = load_dataset("ULM-DS-Lab/SawitMVC", data_files="data/ground_truth.parquet")

import json
from pathlib import Path

tree = json.loads(Path("json/DAMIMAS_A21B_0001.json").read_text(encoding="utf-8-sig"))
gt = tree["summary"]["by_class"]
total = tree["summary"]["total_unique_bunches"]

YOLO Training

yolo detect train data=data.yaml model=yolov8n.pt epochs=100 imgsz=960

Dataset Collection

• Source: Field surveys at DAMIMAS and LONSUM palm oil plantations in Indonesia
• Capture: Smartphone cameras, 4-8 positions per tree
• Annotation: Expert agronomists using multi-view cross-referencing
• Resolution: 960 x 1280 pixels
• Date: February 2026

Citation

@dataset{ulm_sawitmvc_2026,
  title     = {SawitMVC},
  author    = {Fatma Indriani and Setyo Wahyu Saputro and Muhammad Zainal Muttaqin and Alia Rahmi and Triando Hamonangan Saragih and Rahmat Budianoor and Hartoni and Dwi Kartini and Naufal Said},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/ULM-DS-Lab/SawitMVC}
}

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

You may share and adapt the dataset for non-commercial purposes with appropriate attribution. Commercial use is not permitted.