---
license: mit
task_categories:
- image-to-text
- object-detection
- physics-reasoning
tags:
- physics
- vlm
- dataset-curation
- robotics
pretty_name: Canonical Object-Physics Dataset (v3.0)
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: train
    path: metadata.jsonl
---

# Canonical Object-Physics Dataset (v3.0)

This dataset is designed for training and evaluating Vision-Language Models (VLMs) on physical property estimation tasks. It contains 216 high-quality, balanced annotations mapping visual instances to their ground-truth physical properties (mass and friction).

## Dataset Summary

- **Total Samples:** 216 (Cleaned & Balanced)
- **Source Videos:** 5 YouTube kitchen/cooking videos
- **Unique Object Instances:** 162
- **Unique GT Object IDs:** 78
- **Category Diversity:** 7 labels (bottle, bowl, cup, fork, knife, spoon, wine glass)
- **Mass Range:** 0.0019 kg — 1.6084 kg
- **Mass Ratio:** 846.5x (Exceptional diversity for physical reasoning)

## Data Structure

The dataset follows the Hugging Face `imagefolder` format.

- `data/`: Contains all image frames (renamed as `video_XX_frame_XXXXXX.jpg`).
- `metadata.jsonl`: Contains annotations for each image.

### Metadata Fields

| Field | Type | Description |
|---|---|---|
| `file_name` | `string` | Relative path to the image file. |
| `object_instance_id` | `string` | Unique tracking ID for the instance in the video. |
| `gt_object_id` | `string` | ID matching the `physical_ground_truth_final.json`. |
| `physical_properties` | `dict` | Contains `mass_kg`, `mass_range_kg`, and `friction_range`. |
| `object_bbox` | `list` | Bounding box in `[x1, y1, x2, y2]` format. |
| `object_label` | `string` | Text category label exported from YOLO/GT. |
| `annotation_confidence`| `float` | YOLOv8n detection confidence. |

## Stats by Label

| Label | Count |
|---|---|
| cup | 40 |
| bowl | 40 |
| knife | 40 |
| bottle | 40 |
| fork | 40 |
| spoon | 14 |
| wine glass | 2 |

## Quality & Balancing

This dataset has been processed through a multi-stage cleaning pipeline:
1. **Deduplication:** Maximum of 5 temporal frames per unique object instance, selecting for highest detection confidence.
2. **Label Balancing:** Capped at 40 samples per category to prevent majority class bias (standardizing on `bottle`, `bowl`, `cup`, `fork`, `knife`).
3. **Diversity Audit:** Verified that the max/min mass ratio exceeds 3x (Actual: 846x).

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("your-username/phys-vlm-dataset", split="train")
print(dataset[0])
```

## Dataset Generation

Created sequentially on an Apple M2 Air (CPU-only) using YOLOv8n and stable IoU tracking.
Ground-truth properties derived from a canonical household object database.