Datasets:
File size: 2,719 Bytes
389f33c 46b607c 389f33c 46b607c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 |
---
license: mit
task_categories:
- robotics
tags:
- trajectory-prediction
- mouse-control
- computer-control
- quick-draw
- diffusion
size_categories:
- 10K<n<100K
---
# Quick, Draw! Circles - Trajectory Dataset
Dataset for training trajectory prediction models, specifically designed for the [Qwen-DiT-Draw](https://github.com/HusseinLezzaik/qwen-dit-draw) project.
## Dataset Description
This dataset contains chunked trajectory data from the [Quick, Draw!](https://quickdraw.withgoogle.com/data) circle category, formatted for training diffusion-based trajectory prediction models.
### Key Features
- **Variable-length trajectories** with stop signals (GR00T-style)
- **16-point chunks** with (x, y, state) format
- **Loss masking** for handling variable-length final chunks
- **512×512 canvas images** showing drawing progression
## Dataset Statistics
| Metric | Value |
|--------|-------|
| Total samples | 21207 |
| Source circles | 10000 |
| Chunk size | 16 points |
| Canvas size | 512×512 |
| Avg chunks/circle | 2.1 |
## Data Format
Each sample contains:
```python
{
"image": Image, # 512×512 canvas (white for first chunk, partial drawing for rest)
"instruction": str, # "draw a circle"
"trajectory": [[x, y, state], ...], # 16 points, normalized [0, 1]
"mask": [1, 1, ..., 0, 0], # 1=real point, 0=ignore in loss
"is_last": bool, # True if final chunk of trajectory
"n_real_points": int, # Number of real points in this chunk (1-16)
"circle_idx": int, # Source circle index
"chunk_idx": int, # Chunk index within circle
}
```
### State Signal
- `state = 0`: Continue drawing
- `state = 1`: Stroke complete (STOP)
The model learns WHERE to place the stop signal, not a fixed position.
### Loss Masking
For final chunks with fewer than 16 real points:
```
mask = [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0]
↑ real points (count in loss) ↑ ignored
```
## Usage
```python
from datasets import load_dataset
dataset = load_dataset("TESS-Computer/quickdraw-circles")
# Access a sample
sample = dataset["train"][0]
image = sample["image"] # PIL Image
trajectory = sample["trajectory"] # List of [x, y, state]
mask = sample["mask"] # Loss mask
```
## Source
Data sourced from [Google Quick, Draw! Dataset](https://github.com/googlecreativelab/quickdraw-dataset) (circle category only).
## License
MIT License
## Citation
```bibtex
@misc{quickdraw-circles-trajectory,
title={Quick, Draw! Circles Trajectory Dataset},
author={TESS Computer},
year={2025},
url={https://huggingface.co/datasets/TESS-Computer/quickdraw-circles}
}
```
|