Anas-Mohiuddin-Syed/visquant-v1

🔢 VisQuant — The Object Counting Benchmark for Vision-Language Models

LLaVA-7B scores 30%. GPT-4V struggles. Can yours do better?

---

💡 Why VisQuant Exists

Counting is supposedly trivial. You learned it at age 3.

Your frontier VLM didn't.

We tested LLaVA-7B on images with 1–20 everyday objects and it scored 30% overall — worse on hard scenes, barely better than random on expert-level images. VisQuant is a structured, difficulty-graded benchmark with zero label noise (counts are set by code, not annotators) designed to expose exactly where and why vision-language models fail at quantitative reasoning.

Every label in this dataset is mathematically guaranteed correct. No crowdworkers. No ambiguity. No hallucinated annotations.

---

📊 Dataset at a Glance

🖼️ Total Images	67,093
❓ QA Pairs	369,612
🏷️ Object Categories	88 everyday objects
🎯 Count Range	1 – 20 objects per image
📐 Difficulty Levels	Easy → Medium → Hard → Expert
✅ Label Accuracy	100% (programmatically guaranteed)
📄 License	CC BY 4.0

---

🏆 Leaderboard

Submit your model's results — open a discussion with your scores and we'll add you to the board.

Rank	Model	Size	Easy	Medium	Hard	Expert	Overall
🥇	Submit yours!	—	—	—	—	—	—
—	LLaVA-7B	7B	45.0%	28.3%	28.3%	16.7%	29.6%
—	moondream	1.8B	26.7%	18.3%	13.3%	5.0%	15.8%
—	Random chance (est.)	—	~25%	~20%	~15%	~12%	~18%

Metrics: exact-match accuracy on exact_count questions. 240 images tested per model (60/difficulty), temperature=0. LLaVA-13B results coming soon.

---

🎯 Difficulty Breakdown

Easy    ████░░░░░░░░░░░░  7,183 images  — 1–3 objects, simple spatial layouts
Medium  ████████████░░░░  25,857 images — 2–6 objects, mixed arrangements, 1 type
Hard    ████████████░░░░  21,390 images — 3–12 objects, dense scenes, 2–3 types  
Expert  ████████░░░░░░░░  12,663 images — 5–20 objects, multi-type chaos

Easy is a warm-up. Expert is a wall.

---

❓ Question Types

Each image comes with up to 6 question types:

Type	Example	Count
exact_count	"How many feathers are in this image?" → "7"	140,894
threshold	"Are there more than 5 rocks?" → "Yes"	67,093
comparison	"Are there more shells or buttons?" → "Shells"	50,620
total_count	"How many objects are there in total?" → "11"	46,403
spatial	"How are the objects arranged?"	48,129
background	"What surface are the objects on?"	16,473

---

🚀 Quick Start

from datasets import load_dataset

ds = load_dataset("Anas-Mohiuddin-Syed/visquant-v1")

Or load a specific difficulty split:

import json

# Stream easy examples only
with open("data/dataset.jsonl") as f:
    easy = [json.loads(l) for l in f 
            if json.loads(l)[list(json.loads(l).keys())[0]]["difficulty"] == "easy"]

Evaluate your model:

from PIL import Image

sample_id = "visquant_00101"
image = Image.open(f"data/images/{sample_id}.png")

# Ask your VLM the first question
entry  = ds[sample_id]
q      = entry["questions"][0]["q"]   # "How many twigs are in this image?"
gt     = entry["questions"][0]["a"]   # "4"

prediction = your_vlm(image, q)
correct = prediction.strip() == gt

---

🗂️ Data Format

Each record in data/dataset.jsonl:

{
  "visquant_00101": {
    "objects": { "twig": 4 },
    "total_objects": 4,
    "difficulty": "medium",
    "background": "a marble surface",
    "spatial_info": "exactly 4 twigs arranged in a semi-circle on a marble surface",
    "questions": [
      { "type": "exact_count", "q": "How many twigs are in this image?",       "a": "4"   },
      { "type": "threshold",   "q": "Are there more than 3 twigs?",             "a": "Yes" },
      { "type": "spatial",     "q": "How are the objects arranged?",            "a": "Twigs arranged in a semi-circle" },
      { "type": "background",  "q": "What surface are the objects on?",         "a": "Marble surface" }
    ]
  }
}

Images live at data/images/{image_id}.png — 512×512px PNG, transparent-composited sprites on photorealistic backgrounds.

---

🔬 How Images Are Built (No Annotation Needed)

Twemoji SVG sprites  ──►  PIL compositor  ──►  Exact N objects placed
     (88 types)                │                 at computed positions
                               ▼
                    40+ background textures
                    (marble, wood, concrete…)
                               │
                               ▼
                    Labels written by the same
                    code that placed the objects
                    ──► Zero annotation error

Every count label equals the loop variable N that placed those objects. There is no human in the loop, and therefore no human error.

---

📦 What's Inside

visquant-v1/
├── data/
│   ├── dataset.jsonl       # 67,093 records, 369,612 QA pairs
│   └── images/
│       ├── visquant_00101.png
│       ├── visquant_00102.png
│       └── ... (67,093 total, 512×512px)
└── README.md

---

🧩 Object Categories (88 total)

feather · shell · button · leaf · acorn · ring · twig · bead · earring · bracelet · rock · pinecone · sock · hair clip · ribbon · bow · domino · remote control · flower · pebble · marble · coin · candle · eraser · chess piece · apple · banana · cup · dice · USB drive · pencil · stapler · binder clip · plate · mouse · cylinder · glass · puzzle piece · and 50 more…

---

📉 What the Numbers Reveal

Running LLaVA-7B across 240 VisQuant images surfaces a crisp, reproducible pattern:

The subitizing cliff at 3:

1 object  → 75.0% accuracy   ████████████████████████████████████████
2 objects → 81.3% accuracy   ████████████████████████████████████████
─────────────────────────────────── cliff ───────────────────────────
3 objects → 18.4% accuracy   ████████
4 objects → 46.2% accuracy   ███████████████████
5 objects → 25.0% accuracy   ██████████
6 objects → 15.0% accuracy   ██████
7 objects → 47.4% accuracy   ███████████████████  ← grid layout boost
8+ objects → collapses toward 0%
12+ objects → 0.0% (absolute failure)

Systematic undercounting bias: 20% of predictions are off by −1, vs 15.8% off by +1. The model consistently underestimates.

Expert ceiling: 16.7% — barely above a random guess for high-count, multi-type scenes.

VisQuant isolates these failure modes in a controlled, reproducible benchmark so you can measure, not guess how your model handles counting.

---

📋 Reproduce / Extend

The full generation pipeline is available. To regenerate or extend the dataset:

1. Twemoji sprites → rasterized PNG (96×96px)
2. PIL compositor places exactly N sprites per layout type
3. Backgrounds: 40+ PIL-generated textures
4. QA pairs generated rule-based from metadata
5. Images filtered to remove layout overflow (stacked/row >4 objects)

---

🔖 Citation

If you use VisQuant in your research, please cite:

@dataset{visquant_v1_2025,
  title     = {VisQuant: A Programmatic Object Counting Benchmark for Vision-Language Models},
  author    = {Syed, Anas Mohiuddin},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Anas-Mohiuddin-Syed/visquant-v1},
  license   = {CC-BY-4.0},
  note      = {67,093 images · 369,612 QA pairs · 88 object categories}
}

---

📬 Contact & Contributions

• Submit benchmark results: Open a Discussion
• Report issues: Found a bad image? Open an issue in Discussions
• Extend the dataset: PRs welcome — add new object types, layouts, or difficulty tiers

---

_{Built with 🔢 and zero annotation error · CC BY 4.0 · 2025}