harpreetsahota/kubricount-subset

Dataset Card for KubriCount (FiftyOne subset)

kubricount-subset is a FiftyOne image dataset built from a subset of liuchang666/KubriCount, a large-scale synthetic benchmark for multi-grained visual counting. In addition to the ground-truth counting supervision, this FiftyOne version ships open-vocabulary detection predictions from NVIDIA's LocateAnything-3B, SigLIP embeddings, and per-sample counting evaluation fields.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

dataset = load_from_hub("harpreetsahota/kubricount-subset")
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

KubriCount reframes open-world counting as a prompt-following problem across five explicit semantic granularity levels. Each 1024×1024 synthetic scene is produced with a Kubric + Blender rendering pipeline, refined by mask-conditioned image editing, and quality-filtered by a VLM. Every scene comes with pixel-perfect instance masks, 2D boxes, and per-object center points generated directly by the rendering engine (no human annotators).

This FiftyOne dataset is a subset of the full KubriCount release. It is a flat image dataset — one sample per counting query — sampled primarily from level 1 (identity) and level 5 (concept) queries across the train, testA, and testB splits.

• Source dataset: liuchang666/KubriCount
• Paper: Count Anything at Any Granularity (Liu, Wu & Xie, SJTU 2026)
• Project page: verg-avesta.github.io/KubriCount
• Code: Verg-Avesta/KubriCount
• License: Apache-2.0

Counting granularity levels

Level	Granularity	Prompt example	Distractor
L1	Identity	"Count all the dogs."	None
L2	Attribute (size/color)	"Count large cherries."	Small cherries
L3	Category	"Count the cans."	Bags
L4	Instance type	"Count backpack A."	Backpack B
L5	Concept	"Count the lobsters."	Octopuses

This subset focuses on L1 and L5 queries.

---

FiftyOne Dataset Structure

Dataset name: kubricount-subset

Media type: image

Summary

Property	Value
Samples (counting queries)	6,736
Splits	train (5,000), testA (924), testB (812)
Levels present	L1 (5,000), L5 (1,736)
Categories	157

Ground-truth fields

Field	Type	Description
image_id	string	Relative path key matching the source annotation files
split	string	train, testA, or testB
level	int	Counting granularity level
category	string	Target object class to count
count	int	Ground-truth target count
target_points	fo.Keypoints	One keypoint per target object center; len(target_points.keypoints) == count
example_boxes	fo.Detections	2–8 few-shot exemplar boxes for the target class
segmentation	fo.Segmentation	Per-instance segmentation map (mask_path on disk)
negative_category	string	Distractor class (empty for L1)
negative_count	int	Ground-truth distractor count (0 for L1)
negative_points	fo.Keypoints	One keypoint per distractor object
negative_example_boxes	fo.Detections	Few-shot exemplar boxes for the distractor class

Model predictions and evaluation fields

This subset includes open-vocabulary detection predictions generated with NVIDIA LocateAnything-3B through the FiftyOne remote zoo model integration Burhan-Q/fiftyone-locate-anything. The model was run in detect mode using each sample's category as the target class, then evaluated against the ground-truth counts.

Field	Type	Description
locate_anything_pred	fo.Detections	LocateAnything-3B detections for the target category. Each detection carries eval_boxes (tp / fp / fn) marks
predicted_count	int	Number of predicted target detections (len(locate_anything_pred.detections))
abs_error	float	Absolute counting error, abs(predicted_count - count)
eval_boxes_tp / eval_boxes_fp / eval_boxes_fn	int	Per-sample TP / FP / FN box counts from the detection evaluation
siglip_embeddings	VectorField	SigLIP image embedding
siglip_viz	list(float)	2D embedding-visualization coordinates

Note on LocateAnything-3B: the model emits no per-detection confidence scores (fo.Detection.confidence is None), so it is used here as a class-conditioned localizer/counter. Its weights are under the NVIDIA License (non-commercial research only); see the wrapper repo for details.

Example queries

from fiftyone import ViewField as F

# Concept-level test split
view = dataset.match((F("split") == "testB") & (F("level") == 5))

# Where the counter was most wrong
hardest = dataset.sort_by("abs_error", reverse=True)

# Dense scenes
dense = dataset.match(F("count") > 100)

In the App, toggle target_points, example_boxes, segmentation, and locate_anything_pred to compare ground truth against the model's predictions, and color by eval_boxes on the predictions to inspect TP/FP boxes.

---

Citation

@article{liu2026count,
  title   = {Count Anything at Any Granularity},
  author  = {Liu, Chang and Wu, Haoning and Xie, Weidi},
  journal = {arXiv preprint arXiv:2605.10887},
  year    = {2026}
}

License

The source dataset is released under the Apache-2.0 License. LocateAnything-3B model weights are released under the NVIDIA License (non-commercial research only).