kanguyen-vn/M4Seg

M⁴Seg

M⁴Seg is the multi-image pixel-grounded reasoning dataset introduced in PRIMA: Multi-Image Vision-Language Models for Reasoning Segmentation. Each example pairs two or three ordered images with a question, an answer, and object or part masks. Questions cover functional, spatial, open-ended, and numerical reasoning.

PRIMA checkpoint · Paper · Project

Splits

Split	Examples	Grounding annotations
Train	743,463	2,595,754
Test	1,000	3,620

There are exactly two splits. The test split preserves the original benchmark examples. Questions, answers, image order, and mask pixels are preserved from the supplied splits; polygon masks are represented as lossless compressed COCO RLE.

Some original training pair_id values repeat, so example_id provides a unique, stable row identifier. pair_id remains available as source provenance. Test example IDs retain the original pair IDs.

Load with Hugging Face Datasets

pip install datasets pycocotools

from datasets import load_dataset

dataset = load_dataset("kanguyen-vn/M4Seg")
print(dataset["train"].num_rows)  # 743463
print(dataset["test"].num_rows)   # 1000
example = dataset["test"][0]

For streaming access:

dataset = load_dataset("kanguyen-vn/M4Seg", streaming=True)
iterator = iter(dataset["train"])
example = next(iterator)
iterator.close()

The viewer and load_dataset use the Parquet files under data/, with one example per row. The compressed JSON files under annotations/ provide the equivalent native format for the PRIMA training and evaluation loader.

Fields

Field	Description
example_id	Unique example identifier within the release.
pair_id	Original sample identifier; may repeat in training.
dataset_name	Constituent image/annotation source.
type	functional, spatial, open-ended, or counting.
granularity	Annotation level of the source dataset: object or part. Part-source examples may ground objects, parts, or both.
q, a	Original question and answer. Entity codes in answers link phrases to grounding annotations.
images	Ordered list of source image IDs, relative paths, heights, and widths.
groundings	Ordered grounding annotations with image indices, labels, source annotation IDs, answer spans, boxes, and RLE masks.

granularity describes the source annotation family. The part-level sources are ADE20KPart234, PACO-LVIS, and PASCAL-Part; the object-level sources are their counterparts. It is not a per-target label.

Each entry in images contains image_id, path, height, and width. Grounding image_index values start at one and refer to that ordered list. Each grounding contains:

• name: the object or part label.
• annotation_ids: source annotation IDs, normalized to a list of strings. A combined grounding can refer to more than one source annotation.
• object_annotation_id: the source parent-object annotation ID, stored as a string.
• answer_span: original start/end character offsets in a, with an exclusive end offset.
• bbox: [x, y, width, height] computed from the released mask in native image coordinates.
• segmentation: compressed COCO RLE with size: [height, width] and an ASCII counts string.

Groundings are grouped by image index and preserve their within-image order. The native JSON format uses groundings.img_1, groundings.img_2, and subsequent image keys, and retains the original annotation identifier and box representations. It also stores image_paths and image_sizes in image order.

Decode a mask

from pycocotools import mask as mask_utils

grounding = example["groundings"][0]
rle = grounding["segmentation"]
mask = mask_utils.decode({
    "size": rle["size"],
    "counts": rle["counts"].encode("ascii"),
})
# mask is a uint8 array at the original image height and width.

No image download is needed to decode or score a released mask. To inspect it alongside a photograph, resolve the corresponding relative image path under your own dataset root.

Source images

Obtain images from the original projects and follow their access conditions:

M⁴Seg source	Original images
ade20k_part234	ADE20KPart234 image layout from OV-PARTS.
ade20k_objects	ADE20K / Scene Parsing training and validation images.
paco_lvis	COCO2017 images referenced by PACO-LVIS.
coco_objects	COCO2017 training and validation images.
pascal_part, pascal_objects	PASCAL VOC2010 images used by PASCAL-Part.

Use this layout, or create equivalent symlinks to images you already have:

data/
├── coco_2017/
│   ├── train2017/
│   └── val2017/
├── Part_Segm/
│   ├── ADE20KPart234/images/
│   │   ├── training/
│   │   └── validation/
│   └── PASCAL-Part/VOCdevkit/VOC2010/JPEGImages/
└── Semantic_Segm/ade20k_instances/images/
    ├── training/
    └── validation/

Use separate image roots for ADEChallengeData2016 objects and ADE20KPart234 parts: matching filenames can have different JPEG pixels. Follow the preparation guide to create the image layout and verify file checksums. The task masks and native dimensions are included in the release.

Source preparation

See PREPARATION.md for extraction layouts, the image setup/checking tool, the ADE20KPart234 class map, and compact source catalogs for original-format annotations.

Use with PRIMA

Use the PRIMA environment with Python 3.10 and PyTorch 1.13.1 / CUDA 11.7. Download the complete PRIMA source and evaluator, then install its requirements and obtain the native annotations:

huggingface-cli download kanguyen-vn/PRIMA \
  --include "code/*" --local-dir ./prima_release
cd prima_release/code
pip install -r requirements.txt
huggingface-cli download kanguyen-vn/M4Seg \
  --repo-type dataset --local-dir ./data/m4seg \
  --include "annotations/*" "metadata/*" "prepare_images.py" "PREPARATION.md"

The generation driver and metric implementation are included in that source bundle. Follow the image layout in PREPARATION.md, then run:

python evaluate.py \
  --merged_ckpt_path kanguyen-vn/PRIMA \
  --dataset_dir ./data \
  --m4seg_annpath ./data/m4seg/annotations/m4seg_test.json.gz \
  --output_save_path ./outputs/m4seg_test \
  --compute_metrics --text_model sentence-transformers/all-MiniLM-L6-v2

The semantic encoder is sentence-transformers/all-MiniLM-L6-v2, the model used by the paper's metric implementation. This release pins the metric encoder to revision 1110a243fdf4706b3f48f1d95db1a4f5529b4d41. It is downloaded automatically; Java is required for METEOR. The evaluation guide describes all seven metrics and standalone scoring.

The loader reads compressed JSON directly. For further fine-tuning, use the training annotation file with train.py; keep the test split for evaluation. The standard PRIMA training and metric dependencies are required for these commands.

Files

data/train-*.parquet
data/test-*.parquet
annotations/m4seg_train.json.gz
annotations/m4seg_test.json.gz
dataset_summary.json
prepare_images.py
PREPARATION.md
metadata/image_manifest.jsonl.gz
metadata/ade20k_part234_classes.json
metadata/source_catalogs/*/*.json.gz

License

New M⁴Seg annotations are released under CC BY 4.0. Inherited source masks, annotations, and images retain their upstream terms; conversion to RLE does not change them. The original preparation code is licensed under MIT. See LICENSE.md for the license scope, attribution requirements, and full MIT text.

Citation

If you use M⁴Seg, please cite PRIMA:

@inproceedings{nguyen2027prima,
  title     = {PRIMA: Multi-Image Vision-Language Models for Reasoning Segmentation},
  author    = {Nguyen, Kiet A. and Wahed, Muntasir and Juvekar, Adheesh Sunil and
               Li, Xinzhuo and Zhou, Xiaona and Shah, Vedant and Yu, Tianjiao and
               Yanardag, Pinar and Lourentzou, Ismini},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  year      = {2027}
}